sub-agent-review

Pass

Audited by Gen Agent Trust Hub on Sep 18, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data (local code changes and diffs) which may contain malicious instructions designed to influence the agent's behavior.
  • Ingestion points: The skill ingests code changes via sub-agents executing the review-diff skill as described in Step 1.
  • Boundary markers: The prompt template provided in Step 1 for the sub-agent lacks explicit delimiters or instructions to ignore natural language commands embedded within the code diffs.
  • Capability inventory: The skill possesses the capability to modify the local filesystem (Step 3: 'Apply fixes') and execute local shell commands (Step 3: 'Run the project's linting tool').
  • Sanitization: There is no evidence of sanitization or validation of the findings returned by the sub-agent before they are used to justify automated code modifications.
  • [COMMAND_EXECUTION]: The skill includes instructions to autonomously execute local tools.
  • Evidence: Step 3 instructs the agent to 'Run the project's linting tool, if known.' while implementing fixes. While standard for development workflows, this execution is triggered automatically following potentially tainted sub-agent feedback.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 18, 2026, 04:33 AM