Free guide
Cloudflare's security-audit Skill: Have Your Coding Agent Hunt for Security Holes
The open-source skill that seeded Cloudflare's own bug-hunting system. What it does, why no agent gets to check its own work, how to set it up, and the fine print.

The short version
- Cloudflare's open-source security-audit skill (MIT) is trending on GitHub with over 70,000 stars. It's the skill that seeded Cloudflare's own vulnerability discovery harness.
- Its main rule: the agent that checks a finding is never the agent that found it. A fresh verifier tries to disprove every candidate.
- Cloudflare says one run found roughly half of what repeated runs found in total, and runs on the same repo build on each other, so run it more than once.
How it actually works
Why an agent can't grade its own homework
Cloudflare's write-up says an unchecked agent can edit the code it's testing so its own exploit works, then triumphantly report the bug it just created. In their words: "If a Hunter is allowed to grade its own homework, it will confidently validate everything it outputs." So the skill splits finding from checking.
Map the house, then try every door
First, agents map your app like a floor plan: its architecture, trust boundaries and input surfaces, written to architecture.md and a coverage ledger. Then isolated hunters each take a different slice of that map and try to break in, while coverage critics look for rooms nobody checked.
A fresh agent tries to prove it wrong
Every unique candidate goes to a fresh verifier whose job is to disprove it. Each lead ends up as confirmed (a complete source trace), needs_validation (the exact missing fact, no severity) or rejected (disproved). The records are checked against a JSON schema, then fresh agents verify the final claims again before the reports are written.
Run it again
In Cloudflare's test runs, a single run found roughly half of the vulnerabilities that repeated runs found in total. Runs against the same repo are additive: the skill reads its earlier ledgers and findings and goes after the gaps.
How to set it up
1.Have a coding agent that can run sub-agents
The skill needs a coding agent with a model that supports tool use and parallel sub-agents (Claude Code works), plus Node.js for its two validators. Use macOS, Linux or WSL: the validators refuse every input on native Windows (open GitHub issue 53).
node --version2.Install the skill with the Skills CLI
Run this in the project you want to audit. Add --global to install it for your user instead of one project. Run npx skills --help to pick which agent it installs into.
npx skills add https://github.com/cloudflare/security-audit-skill \ --skill security-audit3.Ask for an audit
Start your agent in the codebase and ask in plain words. A direct audit request runs the full six-phase workflow; a security question only uses guidance mode.
security audit this codebase find security vulnerabilities in ./src do a security review, output to ~/audits/my-project4.Read the reports
By default the results go outside your repo, to ~/security-audit-skill/<repo-name>/run-<N>. Open REPORT.md first, then FINDINGS-DETAIL.md for confirmed findings and NEEDS-VALIDATION.md for the leads it couldn't settle. The machine-readable version is findings.json.
5.Fix, then run it again
Review each fix yourself before you ship it, then run the audit again on the same repo. It builds on the earlier run and targets what it didn't cover.
The fine print
- Free and open source (MIT, Cloudflare). The skill costs nothing; the model tokens your agent spends are what you pay.
- It isn't cheap to run. One outside tester (three runs per method, one test codebase, and they disclose they build a competing tool) measured a median of $29.95 per quick-profile run vs $2.06 for a single-agent review, with the same recall (GitHub issue 20).
- It describes fixes; it doesn't change your source. If you have your agent apply fixes afterwards, check them: one user reported that more than a dozen fixes also removed features they had built on purpose (GitHub issue 88).
- It's source-first and defensive. It doesn't probe live sites or external services, and it only runs your app's code inside an OS-enforced sandbox with no network. Without one, those leads stay needs_validation.
- It's not new: the repo was created June 18, 2026 and the last code change was September 14. It's trending again after Cloudflare's developer account posted it on November 4.
- This is the skill, not Cloudflare's full harness. The fleet-wide harness from the blog post hasn't been released.
FAQ
Is Cloudflare's security-audit skill free?
Yes. It's open source under the MIT license. You still pay for the model your coding agent uses, and a full audit can use a lot of tokens.
Does it work with Claude Code?
Yes. It's agent-neutral: it works with coding agents that support tool use and parallel sub-agents, and the Skills CLI can install it into Claude Code. It's made by Cloudflare, not Anthropic.
Will it attack my live website?
No. It reads your source code and only runs your code inside a local sandbox with no network. Its rules forbid probing deployed endpoints, external services or production data.
Does one audit catch everything?
No. Cloudflare says a single run found roughly half of what repeated runs found in total, and they don't claim a false-negative rate. Run it more than once; each run builds on the last.
What do confirmed, needs validation and rejected mean?
Confirmed has a complete source trace and an observed result. Needs validation is a credible lead with one exact fact still missing, and no severity. Rejected is a candidate a fresh verifier disproved.
Sources
- security-audit-skill on GitHub
- Cloudflare blog: Build your own vulnerability harness
- @CloudflareDev post (Nov 4, 2026)
- Issue 20: an outside cost and recall test
Facts checked 2026-11-12 against the security-audit-skill README, SKILL.md, license, open GitHub issues 20, 53 and 88, Cloudflare's blog post and the GitHub API (70,592 stars).
Want help building with tools like this?
Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.
Join Shipping SkoolMore free guides
How ChatGPT Wallet Pays Without Seeing Your Card (and How to Set It Up)
OpenAI's new ChatGPT Wallet lets your dot pay for things on your behalf, and neither the agent nor the store ever sees your real card number. Here's how the stand-in card works and how to set it up.
Claude Code Custom Themes: Animated Backdrops and Your Own Colors
The Claude Code desktop app can now wear a theme: your own colors and an animated scene behind the chat. Here is how a theme works, two ways to get one today, and why you should treat a theme like code.
Astryx: Give Your AI Agent Meta's Design System So Every Screen Matches
Meta's open-source design system ships 150+ React components, seven themes and a CLI your coding agent reads from. What it is, how it works, how to set it up today, and the fine print.