Cloudflare's Audit Skill Bans Agents From Grading Their Own Work

Cloudflare open-sourced a security-audit skill that turns a coding agent into a multi-phase auditor. Its core rule: the agent that finds a bug is never the agent that confirms it.

The stakes go beyond vulnerability scanning. Any team running agents unsupervised has hit the same failure: an agent reports a task done without actually checking it worked.

The skill runs six phases in sequence: recon, hunt, validate, report, structured output, and independent verification. The hunt phase sends parallel agents across separate vulnerability categories, then a different set of agents tries to disprove each finding.

Findings come out as JSON, checked against a report-schema.json file by a zero-dependency Node validator. The skill is agent-agnostic, needing only a model with tool use and the ability to run parallel subagents, per the repo.

The adversarial split is the part worth copying elsewhere. Adversarial verification means a separate process checks a claim instead of trusting the one who made it. It applies anywhere an agent self-reports success: code review, data pipelines, test suites. For context, a related tool, GitHub's Agentic Autofix, already runs an explore-fix-verify loop before opening a PR. An independent verification step is a small addition from there.

The repo doesn't say how verifying agents get picked, whether it's random assignment or a deliberate cut from the original hunt's context. That's what decides whether the adversarial step is real independence or a second look with the same blind spots.