apply-code-review
Pass
Audited by Gen Agent Trust Hub on Nov 6, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data, specifically code review reports and development plans, which directly influence the agent's code modification logic and verification steps. While the instructions emphasize that reports are "claims to be verified" and not "automatic instructions," the lack of strict boundary markers or sanitization for these inputs creates a surface for indirect prompt injection.
- Ingestion points: Section 1 and 2 of
SKILL.mddescribe reading user-provided review reports, development plans, and historical implementation records. - Boundary markers: None. The skill relies on natural language instructions for the agent to distinguish between data and instructions.
- Capability inventory: The skill performs file system writes (fixing code, updating records in
*.records/M*.md) and executes shell commands for verification (running tests and verification steps defined in the plan). - Sanitization: None. The skill executes "actual commands/steps" from the plan without explicit validation or escaping of the content.
Audit Metadata