adversarial-review

Adversarial AI code review lenses challenging assumptions, testing edge cases, and verifying AI-generated code.

SKILL.md

GitHub

Adversarial Review Skill

Core Directives

  1. Treat Code as Hostile: Assume every external input (headers, query params, cookies, payloads, queue messages, LLM outputs) is attacker-controlled.
  2. Trace Sinks: Follow untrusted data to dangerous interpreters (SQL, shell, DOM, templates, filesystem).
  3. Challenge Test Sufficiency: Verify whether existing tests actually prove correctness or merely test happy paths.
  4. Agent Output Verification: Specifically inspect code generated by AI coding agents for hallucinated imports, omitted error handling, and subtly broken edge cases.