Cloudflare Security Audit Skill: How the 6 Phases Work

A skill that staffs the audit

Turn your coding agent into a team of security auditors with the open-source skill that seeded Cloudflare’s vulnerability discovery harness.

Cloudflare Security Audit Skill is a coding-agent skill for multi-phase security audits. Cloudflare Developers published the source and the instructions in the official repository on github.com, and described it as the skill that seeded the company’s vulnerability discovery harness.

The problem it targets is a familiar one after a wave of vulnerability disclosures: you want a closer look at your own code, and a single agent pass tends to dump unverified guesses. This skill structures that look so later phases can challenge and check what earlier phases propose.

Six phases, then a verdict

Report what’s confirmed, rejected, or still needs validation.

The published approach is a fixed sequence. The agent maps the codebase, hunts for vulnerabilities, challenges each candidate, records the findings, verifies those findings, and only then reports. Verification is independent of the hunt, and the output is meant to be machine-readable so another tool or a later review can consume it.

That last split is the useful part of the report. A candidate can come back confirmed, rejected, or still in need of validation. You are not handed a flat list that treats every suspicion as a bug.

• Map

  The agent builds a picture of the codebase before it starts hunting.

• Hunt and challenge

  It looks for vulnerabilities, then challenges each candidate instead of accepting the first hit.

• Record, verify, report

  Findings are written down, checked, and labeled confirmed, rejected, or still needing validation.

Where it fits in a review

A practical run starts from code you already have. You load the skill into the coding agent you use, point that agent at the repository under review, and let the six phases finish before you treat anything as a defect. Because the findings are machine-readable, the report can feed a tracker, a follow-up agent pass, or a human reviewer who only wants the confirmed set.

The same sequence works as a scheduled look at a service you maintain, or as a one-off pass after you hear about a class of bugs and want to see whether your tree matches it. The skill does not replace reading the items marked as still needing validation. Those stay open until something else checks them.

Compared with asking an agent for a one-shot vulnerability list, the difference is the middle of the pipeline. Challenging, recording, and verifying are separate jobs, and the report is allowed to reject a candidate. You trade a faster dump of suspicions for a smaller set of findings that have already been argued with.