review-verification-protocol
Pass
Audited by Gen Agent Trust Hub on Oct 22, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: The skill consists of instructional guidelines for the agent to follow during code reviews. It does not contain executable code, remote dependencies, or persistence mechanisms.
- [INDIRECT_PROMPT_INJECTION]: The skill processes external data (code and diffs) which is a potential attack surface, but it implements specific defensive instructions to mitigate this risk.
- Ingestion points: The protocol processes code findings, diff hunks, and structured reports provided as input during a code review task in SKILL.md.
- Boundary markers: The 'Anti-confabulation gate' requires the agent to echo the exact artifact (file, line, and code snippet) directly from the parsed source before issuing a verdict, which prevents the agent from being misled by injected instructions in the context.
- Capability inventory: The skill does not define any subprocess calls, file writing capabilities, or network operations.
- Sanitization: It enforces a multi-step verification process (Anchor, Evidence, Severity, and Format) to ensure all findings are grounded in objective artifacts rather than contextually primed inferences.
Audit Metadata