Cloudflare's Security Audit Skill: The Six-Blank Finding Test
The skill reports a vulnerability only when six blanks are filled, sends every candidate to a refuting verifier, and refuses to run code without a sandbox.

Ask a coding agent to review a repo for security and you get twenty bullets: add rate limiting, set CSP headers, rotate secrets. Some matter. Most are not reachable from anything an attacker controls, and the list never says which.
Cloudflare's security-audit skill is written against that habit, and the rule it leans on is portable — you can run it by hand on a report from any tool.
The six blanks a finding has to fill
The skill makes Claude Code report a vulnerability only when it can name all six of these: the lower-trust principal, the accepted input or action, the intended control, the crossed boundary, the affected principal or resource, and the concrete observed or owner-observable result.
A finding that passes reads like: anonymous visitor, ?redirect= query param, an allowlist of redirect hosts, leaves the app origin, a logged-in user's session, a phishing redirect with a valid session. One that fails reads like: "an attacker" with no entry point, no named input, "best practice says", no boundary, the developer's own machine, "could be risky."
The instruction is explicit that a gap alone is not a finding — do not elevate a missing best practice, guessed deployment behavior, a generic parser crash or self-impact into a security finding. When one layer already stops the attack, the README calls the missing next layer a hardening note. Severity follows the same logic, with a one-line tiebreak: if you cannot state the concrete damage, the severity is lower than it feels.
A second agent tries to refute it
Findings are not graded by the agent that produced them. Every candidate goes to a fresh verifier whose prompt starts: "You did not write this candidate. Try to refute it from repository source and bounded local evidence." Each record ends as confirmed, needs_validation or rejected, and two zero-dependency Node validators check findings.json and the coverage ledger before the run may call itself complete.
It refuses to run your code without a sandbox
Execution requires an OS-enforced sandbox: no external network, an allowlisted environment, a read-only target, scratch-only writes and hard resource limits. Miss one control and the skill stops rather than degrading quietly — it reports the missing sandbox capability as a needs-validation blocker and gives a validation plan you can run yourself. Probing deployed endpoints, shared infrastructure or other users' data is forbidden outright.
Loading the skill also does not start an audit. Its operating-modes rule says the skill is guidance by default, and that loading it does not authorize the complete audit workflow or file creation; a security question gets a normal answer with no files written.
Scale, and what to copy
Per the Skillselion directory, which published this breakdown and discloses that it runs an independent catalog of Claude Code, Codex and Cursor extensions, the skill had 118,132 installs in its catalog on 20 November 2026 and the cloudflare/security-audit-skill repo carried about 76,000 GitHub stars, with the newest commit c1c8a8c dated 14 September 2026. The folder ships 20 files: SKILL.md, 14 companion Markdown files, a report schema, two Node validators and their two tests.
The transferable part costs nothing to adopt. Take the last security report any tool handed you, pick one item, and try to fill the six blanks from it. Items that only fill the right-hand column are hardening notes — file them that way, and the real findings stop drowning in them.
More from DangMua