Most AI security skills ask a model to review your code and report what it finds, which is a recipe for confident false positives. Cloudflare’s security-audit skill is built around the opposite assumption: that a single agent’s opinion isn’t trustworthy, so every claim has to survive an agent that’s trying to disprove it. It’s the skill that seeded Cloudflare’s own vulnerability discovery harness, now open source and usable on any single repository.
It lives at cloudflare/security-audit-skill under the MIT licence, and it’s already one of the most-starred repos in this niche, with over 54,000 stars at the time of writing. I’ve read the README and the file list rather than running a full audit, so this is an overview of the design, not a record of findings from my own projects.
Where it came from
The README says this is the single-repo starting point that evolved into a multi-stage, fleet-wide system at Cloudflare, described in their post “Build your own vulnerability harness”. If you want the thinking behind it, that’s the place to read. This repo is the portable piece you can drop into a coding agent today.
The six phases
- Reconnaissance: map the architecture, trust boundaries, input surfaces and prior evidence, recorded in
architecture.mdand acoverage-ledger.json. - Coverage-led hunting: assign isolated hunters to units from the ledger, record what each one checked, and use coverage critics to find gaps.
- Candidate validation: give every unique candidate to a fresh verifier whose job is to try to disprove it.
- Structured output: write
confirmed,needs_validationandrejectedrecords tofindings.json, validated against a JSON schema. - Independent record verification: fresh agents check the final source claims, and material replacements get another independent verifier.
- Target-neutral reporting: derive
REPORT.md,FINDINGS-DETAIL.mdandNEEDS-VALIDATION.mdfrom the verified records.
The verdicts are deliberately strict. confirmed needs a complete source trace and a bounded observed result. needs_validation records the exact unresolved fact and carries no severity. rejected records a disproved candidate. Two small zero-dependency Node.js validators check the findings file and the coverage ledger so that the output stays machine-readable.
Design principles worth stealing
- The agent that checks a finding is never the one that found it.
- Severity requires impact, based on likelihood times impact rather than deviation from a checklist.
- Defence-in-depth gaps aren’t vulnerabilities. If one layer stops the attack, a missing second layer is a hardening note.
- Run it more than once. The authors say a single run found roughly half of the vulnerabilities that repeated runs found in total, and repeated runs are additive because the skill reuses earlier ledgers and findings to target gaps.
What it covers
Beyond the core attack classes, the skill ships separate hunting guides depending on what your target is: memory safety and binaries, AI and LLM apps (prompt injection, agent tools, output handling), HTTP protocol and authentication, client-side code, supply chain and release pipelines, cloud and deployment config, RPC and messaging, resource exhaustion, tenant data isolation and lifecycle, and desktop, mobile and local IPC. Recon decides which ones apply, so a small web app doesn’t get a kernel-hunting prompt.
How to install it
It installs with the Skills CLI:
npx skills add https://github.com/cloudflare/security-audit-skill
--skill security-audit
Add --global for a user-level install, and run npx skills --help for agent-selection options. It’s written for coding agents in general, not just Claude Code.
How to run it
Start your agent in the codebase and ask for an audit in plain language, for example “security audit this codebase”, “find security vulnerabilities in ./src” or “do a security review, output to ~/audits/my-project”. A direct audit or pen-test request runs full audit mode. Security questions and focused vulnerability work use a lighter guidance mode unless you ask for report artefacts.
In full mode, output goes to ~/security-audit-skill/<repo-name>/run-<N> unless you choose a directory. It only writes inside the target repo if you explicitly pick a directory that version control ignores, so your working tree stays clean.
Requirements and cost
- An agent with tool use and parallel sub-agents. The workflow depends on spawning many isolated agents.
- Node.js, for the two validators.
- An OS-enforced sandbox if you want it to build or run target code. It needs to disable external networking, use a sanitised allowlisted environment, enforce resource limits and allow writes only to scratch paths. Without that, the skill keeps the lead as
needs_validationinstead of executing anything from the target.
The README doesn’t give token or cost figures. A six-phase audit with many independent agents and repeat runs is likely to use far more of your quota than a single-prompt review, so start on a small repo and watch your usage.
Things to know
- Audit only code you own or are authorised to test, especially if you let it execute anything.
- Your code goes to your model provider during the audit, so check your policy for private or client repos.
- Don’t skip the sandbox. Running a target’s build or tests is running untrusted code.
- Expect the review to stay human. Verified findings are far better than raw model output, but you still decide what’s real and what to fix.
Where it fits
This is the heavyweight option among the Claude-compatible security skills I’ve covered. The OWASP skill is a light standards reference, vbsec is a fast scan for common issues in AI-written code, and Ghost Security Skills wraps real scanners. Cloudflare’s skill is the one for a deliberate, multi-run audit where you want findings that have been independently challenged before you read them.

Leave a Reply
You must be logged in to post a comment.