Cloudflare's Security Audit Skill Is Trending on GitHub, and One Run Finds Roughly Half the Bugs Repeat Runs Do

Cloudflare's security-audit-skill gained 2,979 GitHub stars in a day, topping the platform's trending list on Oct. 22. The repository, published by the company under an MIT licence, turns a coding agent into a multi-phase security auditor that writes machine-readable findings.

Stars measure attention, not deployment. The README is the only first-party description, and its own efficiency figure is Cloudflare's, from its own test runs.

What the skill does

The README describes "a coding-agent skill for multi-phase security audits with independently verified, machine-readable findings." It says the skill "seeded Cloudflare's vulnerability discovery harness," which it calls the single-repo starting point for a larger system.

It installs with npx skills add https://github.com/cloudflare/security-audit-skill --skill security-audit. It needs a coding agent with tool use and parallel sub-agents, plus Node.js for its validators. No specific agent is named.

Phase	What happens	Output
1. Reconnaissance	Maps architecture, trust boundaries, input surfaces	architecture.md, coverage-ledger.json
2. Hunting	Isolated hunters per ledger unit, plus coverage critics	Candidates
3. Validation	A fresh verifier tries to disprove each candidate	Verdicts
4. Structured output	Schema-checked records	findings.json
5. Verification	Fresh agents check the final source claims	Checked claims
6. Reporting	Human-readable reports	REPORT.md and two detail files

Each finding lands as confirmed, needs_validation or rejected. A separate write-up from Nicolas Deville's notes counts 10 domain playbooks, among them memory safety, supply chain, cloud deployment and AI prompt injection, plus two dependency-free JavaScript validators.

The efficiency claim

The README's own number is modest. "In our test runs, a single run found roughly half of the vulnerabilities that repeated runs found in total." That is a vendor-supplied result, and no independent party is named as having reproduced it. The practical reading is that one pass leaves about half the findable bugs unfound.

Deville's page attributes larger figures to Cloudflare's internal system: 20,799 raw candidates narrowed to 7,245 actionable findings, and a validation rejection rate that fell from 40% to 11%. The page itself says the seed relationship is a claim it cannot verify, and those numbers describe the internal harness, not what this repository will do on a given codebase.

Safety limits in the README

The skill does not run untrusted code freely. For target-controlled builds and tests the README requires "An OS-enforced sandbox" with networking disabled. Without one, it "keeps the lead as needs_validation instead of executing target code."

Full audit output goes to ~/security-audit-skill/<repo-name>/run-<N> by default. Writes land inside the audited repo only if the user picks a directory that version control ignores.

What the star count does not say

The README page showed 38.6k stars, 2.4k forks, 30 open issues and 50 pull requests when fetched on Oct. 22, with 14 commits on main. Deville's notes, dated Sept. 22, listed 19,454 stars, so the count is about twice as high a month later. Both figures are from pages read this week, and neither says how many teams run it.

The repository sits in a week of AI-found vulnerabilities. The Terminal has covered OpenAI's claim that AI found the PHP flaws before attackers, and the Trail of Bits regression hunt.

The repository lists security-ai-research@cloudflare.com as its contact. The next sign of real use would be public findings credited to the skill, and none are linked from the README.