plan-review-cdx

Pass

Audited by Gen Agent Trust Hub on May 5, 2026

Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection. It reads implementation plans from the user context or local files and provides this untrusted content to reviewer subagents. A malicious plan could include instructions intended to bypass the review logic or force specific "corrections."
  • Ingestion points: The active plan context or file paths provided as arguments in SKILL.md (Step 0).
  • Boundary markers: Uses markdown delimiters and horizontal rules to separate reviews, but lacks explicit instructions for subagents to ignore instructions embedded within the ingested plan content.
  • Capability inventory: The agent has file-write permissions to modify the plan, can spawn subagents, and can execute shell commands (cp, rm).
  • Sanitization: No sanitization or validation of the plan content is performed before it is processed by the agents.
  • [COMMAND_EXECUTION]: The skill employs shell commands (cp, rm) for managing plan snapshots and cleaning up temporary files in /tmp/. While these are used for standard workflow state management, they process paths associated with the user-provided plan file.
Audit Metadata
Risk Level
SAFE
Analyzed
May 5, 2026, 05:36 PM