wk-adversarial-review

Pass

Audited by Gen Agent Trust Hub on Aug 28, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONCOMMAND_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from repository diffs and pull request metadata, exposing it to subagent reasoning. \n
  • Ingestion points: Data enters the context through git diff, git log, and gh pr view (specifically PR title and body content). \n
  • Boundary markers: No explicit delimiters or sanitization instructions are specified for the interpolation of untrusted data into prompts. \n
  • Capability inventory: The skill utilizes Bash (restricted commands), Agent (subagent spawning), Read, Glob, and filesystem write/execute capabilities. \n
  • Sanitization: No explicit filtering or sanitization of external content is documented before processing. \n- [DYNAMIC_EXECUTION]: The skill generates and executes scripts in a .review-playground/ directory to validate runtime-behavior claims found during the review process. This involves executing code derived from the untrusted pull request in the local environment. \n- [COMMAND_EXECUTION]: The skill executes various shell commands via the Bash tool for repository interaction and analysis. These calls are restricted at the platform level to specific patterns like git diff:* and gh pr view:*, limiting the risk of arbitrary command injection.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 28, 2026, 12:25 AM
Security Audit — agent-trust-hub — wk-adversarial-review