plan-review
Pass
Audited by Gen Agent Trust Hub on May 6, 2026
Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [PROMPT_INJECTION]: The skill reads and processes untrusted data from implementation plans and source code files. This presents a potential surface for indirect prompt injection, where a malicious plan could contain instructions aimed at influencing the agent's review verdict. The risk is mitigated by the skill's mandatory triple-round review and adversarial simulation requirements.
- Ingestion points: plan.md, phase-*.md, and existing codebase files read via grep or glob.
- Boundary markers: The skill does not explicitly use delimiters to isolate untrusted plan content during processing.
- Capability inventory: The skill can spawn sub-agents (code-reviewer), execute local Python scripts (code_graph), and create workspace tasks.
- Sanitization: No specific content sanitization or validation is mentioned before the analysis rounds.
- [COMMAND_EXECUTION]: The skill executes a project-specific Python script located at .claude/scripts/code_graph to analyze code dependencies and blast radius. This is a functional requirement for the skill's mapping and verification tasks.
Audit Metadata