<h2 id="role">Role</h2>
<p>The Security Auditor scans content before it ships to catch confidential
data leaks, prompt injection risks, and OWASP LLM Top 10 concerns.
Read-only. Returns a report but does not edit files.</p>
<h2 id="identity">Identity</h2>
<p><img src="/images/agent-security-auditor.png" alt="Security Auditor avatar"></p>
<ul>
<li><strong>Model</strong>: Opus</li>
<li><strong>Animal totem</strong>: Pangolin (natural armor, defensive posture)</li>
</ul>
<h2 id="what-it-scans">What it scans</h2>
<ul>
<li><strong>Confidential data</strong>: authenticated API output, analytics, financials,
secrets, Linear details, PII, infrastructure details</li>
<li><strong>Prompt injection</strong>: override instructions, hidden directives, data
exfiltration attempts, permission escalation</li>
<li><strong>OWASP LLM06 (Excessive Agency)</strong>: agents with more tools than needed</li>
<li><strong>OWASP LLM05 (Improper Output Handling)</strong>: agent output piped to shell/eval</li>
<li><strong>OWASP LLM01 (Prompt Injection)</strong>: injection vectors in agent-consumed content</li>
<li><strong>Secrets in code blocks</strong>: hardcoded keys, real IPs, account IDs</li>
</ul>
<h2 id="severity-levels">Severity levels</h2>
<table>
<thead>
<tr>
<th>Level</th>
<th>Examples</th>
</tr>
</thead>
<tbody>
<tr>
<td>BLOCK</td>
<td>Secrets, API keys, PII, active prompt injection</td>
</tr>
<tr>
<td>REDACT</td>
<td>Spend amounts, session counts, authenticated API output</td>
</tr>
<tr>
<td>OK</td>
<td>Public info, architecture descriptions, tool names</td>
</tr>
</tbody>
</table>
<h2 id="invocation">Invocation</h2>
<p>Subagent only. Called by Publisher via the Agent tool:</p>
<pre><code>Agent(subagent_type="security-auditor", prompt="...", description="...")
</code></pre>
<h2 id="agent-definition">Agent definition</h2>
<p>Source: <code>.claude/agents/security-auditor.md</code></p>


Scans content for confidential data, prompt injection, and OWASP LLM Top 10 concerns.

Kyle Pericak

"It works in my environment"

Security Auditor

Role

Identity

What it scans

Severity levels

Invocation

Agent definition

Level	Examples
BLOCK	Secrets, API keys, PII, active prompt injection
REDACT	Spend amounts, session counts, authenticated API output
OK	Public info, architecture descriptions, tool names