Disclosure: we run Skillselion, an independent directory of Claude Code, Codex and Cursor extensions. We have no tie to Anthropic, Cloudflare or GitHub. This post is adapted from our security-audit skill deep dive, and we re-read the upstream files on 20 November 2026 before quoting them.
Short answer: Cloudflare's security-audit skill makes Claude Code report a vulnerability only when it can name the lower-trust actor, the boundary they crossed and the concrete result. A missing best practice goes down as a hardening note. Every candidate then goes to a second agent that tries to refute it, and the skill refuses to run your code without a full sandbox.
Why do agent security reviews read like checklists?
Ask a coding agent to "review this repo for security" and you often 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 does not tell you which.
Cloudflare's skill is written against that habit. On 20 November 2026 it had 118,132 installs in the Skillselion catalog (skills.sh registry data), and the cloudflare/security-audit-skill repo had about 76,000 GitHub stars. The newest commit is c1c8a8c, from 14 September 2026. The skill folder ships 20 files: SKILL.md, 14 companion Markdown files, a report schema, two Node validators and their two tests.
What is the finding test?
This is the rule the rest of SKILL.md leans on:
For every candidate, name the lower-trust principal, accepted input or action, intended control, crossed boundary, affected principal or resource, and concrete observed or owner-observable result. Do not elevate a missing best practice, guessed deployment behavior, generic parser crash, or self-impact into a security finding.
Source: skills/security-audit/SKILL.md, "Require a boundary and result"
You can run this test by hand on any report, from any tool. Take one item and fill in six blanks:
| Blank | Example that passes | Example that fails |
|---|---|---|
| Lower-trust principal | anonymous visitor | "an attacker" with no entry point |
| Accepted input or action |
?redirect= query param |
none named |
| Intended control | allowlist of redirect hosts | "best practice says" |
| Crossed boundary | leaves the app origin | none |
| Affected principal or resource | logged-in user's session | the developer's own machine |
| Concrete result | phishing redirect with a valid session | "could be risky" |
If an item can only fill the blanks the way the right column does, it is not a finding. The README names the most common case: when one layer already stops the attack, the missing next layer "is a hardening note." The anti-pattern list at the end of SKILL.md opens with "Checklist deviations presented as vulnerabilities."
Severity follows the same logic. Only confirmed records get one, and the tiebreak is one line:
If you cannot state the concrete damage, the severity is lower than it feels.
Source: skills/security-audit/SKILL.md, "Separate priority from certainty"
Who checks the agent's findings?
Not the agent that found them. In Phase 3 every candidate goes to a fresh verifier whose prompt (in VALIDATION-AND-REPORTING.md, "Candidate-verifier prompt") starts with "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. Two zero-dependency Node validators check findings.json and the coverage ledger before the run may call itself complete.
Will it execute my code?
Only inside 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 it stops:
If every control cannot be enforced, do not execute target code: report the missing sandbox capability as a needs-validation blocker and give a safe validation plan.
Source: skills/security-audit/SKILL.md, "Universal execution safety"
If your session cannot enforce that sandbox, expect a source-only audit with some leads parked as needs_validation, plus a plan you can run yourself. The skill also forbids probing deployed endpoints, shared infrastructure or other users' data.
Does loading the skill start a full audit?
No. The first rule under "Operating modes" says the skill "is guidance by default. Loading it does not authorize the complete audit workflow or file creation." A security question gets a normal answer with no files written. The six-phase run starts only when you ask to audit or pen-test the codebase, ask for a full or end-to-end review, or ask for report artifacts. If your wording could go either way, the agent asks you one question first.
How much does one run catch?
Cloudflare answers that in its own README:
In our test runs, a single run found roughly half of the vulnerabilities that repeated runs found in total.
Source: README.md, "Design principles"
The workflow is built around that number. Each run writes a coverage ledger; the next run reads earlier ledgers, re-checks changed source and aims at the gaps. You can cap spend as a maximum number of agent calls. The skill reserves its reconnaissance, critic and verifier calls first, and if the budget cannot cover that minimum it launches nothing and asks for a larger budget, a narrower scope or a different profile.
How does it compare with checklist review skills?
They answer different questions. The ECC security-review skill (20,890 installs on 20 November 2026) walks ten numbered categories, from secrets management to dependency security, and ends in a pre-deployment checklist. It tells you what is missing. Cloudflare's skill tells you what someone can do to you.
Other options in the same space, with installs from the catalog on 20 November 2026:
- Sentry's security-review skill, 24,181 installs, and its sibling GitHub Actions security review skill, 5,113 installs, for workflow files.
- The Firebase Security Rules Auditor skill, 151,361 installs, if your risk sits in Firestore or Storage rules.
- The awesome-copilot security-review skill, 5,904 installs.
Should you install it?
Install it when you want findings you would actually fix and you accept a multi-agent run with a verifier per candidate. The README asks for a model with tool use and parallel sub-agents, plus Node.js for the validators. For a fast pre-release pass, a checklist skill is the cheaper tool.
npx skills add https://github.com/cloudflare/security-audit-skill --skill security-audit
Then open Claude Code in the repo and ask it to security audit this codebase. New to skills? Our page on installing a Claude skill covers the folders.
Where to go next
- The Cloudflare security-audit listing has the file panel, FAQ and the full deep dive.
- Compare options in our guide to the best security skills for Claude Code, the ranked security skills page or the security skills category.
- Every skill Cloudflare publishes is on the Cloudflare publisher page.
- How many skills carry any audit at all: our Skill Security Census.
- More file-level reviews like this one: the Skillselion deep-dive index.
Try the six-blank test on the last security report you received. How many items survive it?
Skillselion is a live catalog, refreshed daily from skills.sh, GitHub and MCP registries, ranked by real installs.
Top comments (0)