# Adversarial Review

> Run the adversarial review protocol on a review-gated (security-tradeoff) gap for recurve — try to BREAK the change, never to confirm it

- Skill: `bordumb/adversarial-review` (Agent Skill)
- Install (CLI): `npx skillmds@latest add bordumb/adversarial-review`
- Raw SKILL.md: https://api.skillmd.com/api/skills/bordumb/adversarial-review/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Security
- Author: bordumb (https://skillmd.com/u/bordumb)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/bordumb/adversarial-review

---


# Adversarial review

You were invoked as the INDEPENDENT reviewer of a review-gated gap. You must
not be the agent that implemented the change.

1. `recurve review <ID>` prints the brief: what is now accepted, which
   attacks to attempt.
2. Follow `.recurve/REVIEW.md`: enumerate the accept-delta, invent new attacks beyond
   the existing probes, attack any corroboration source's trust assumption,
   re-read the original fail-closed rationale.
3. Verdict is binary: "broken, here's the input" or "could not break it, and
   here is everything I tried." Every attack you tried must become a new
   probe or trap fixture before anyone promotes.
4. Promotion is the human's call via `recurve adjudicate <ID>` — you report,
   you do not promote.