AI / LLM safety

Why this agent matters

Prompt injection is number one in the OWASP Top 10 for LLM Applications. Once a model can call tools, injected text becomes injected actions.

Traditional scanners don't model this at all. This agent looks at where untrusted text enters a prompt and where model output is executed or trusted.

What it checks

• Untrusted text (user input, fetched pages, files) concatenated into prompts
• Model output executed, or passed into tool, shell or SQL calls
• Model output trusted without validation
• Agent tools broader than the task needs

When it marks a finding as a likely false positive

• Trusted instructions are separated from untrusted content
• Output is validated before it drives an action

The fix it suggests

• Separating instructions from untrusted content
• Validating and constraining model output
• Scoping agent tools narrowly

Scope and limits

This agent validates findings reported by Semgrep, Gitleaks and Trivy; it doesn't hunt for new bugs on its own. A finding reaches it when its CWE or rule/path keywords match this vulnerability class. It sees only the minimal code around the finding, with secrets redacted, and returns a verdict with a confidence score, a plain-English explanation and a suggested patch.

Common questions

Related agents

• Network & transport security — Checks that data in transit is encrypted and certificates are actually verified.
• Privilege & file permissions — Checks file permissions, temp files, symlinks and unnecessary root.
• Mobile client security — Reviews Android and iOS code for on-device storage, transport and component exposure.

See all 20 review agents → · Agent setup reviewer · How the pipeline works