Cloudflare's Security-Audit Skill Hits 33,660 Stars on One Unverified Claim: One Run Finds Half

Cloudflare's open-source security-audit-skill reached 33,660 GitHub stars by Tuesday, October 20, five weeks after its last commit, with a README that publishes one performance figure and no cost estimate.

The repository, cloudflare/security-audit-skill, was created on June 18, 2026 and is MIT-licensed. Its last push was on September 14, according to GitHub's API. It has 2,112 forks and 70 open issues. SmartScope, an independent analysis blog, counted 25,391 stars on October 7, so the repo has added roughly 8,300 since.

What it does, phase by phase

The README describes a skill for coding agents that audits a codebase in six phases. It requires "a coding agent with a model that supports tool use and parallel sub-agents" and Node.js for two validators. It names no specific agent.

Phase	What runs	Output
1. Reconnaissance	Maps architecture and trust boundaries	architecture.md, coverage-ledger.json
2. Hunting	Isolated hunters, plus coverage critics	Candidate findings
3. Validation	A fresh agent tries to disprove each candidate	Verdicts
4. Structured output	Records sorted three ways	findings.json
5. Record verification	New agents check source claims	Corrected records
6. Reporting	Derives reports	REPORT.md, FINDINGS-DETAIL.md, NEEDS-VALIDATION.md

The three verdicts are confirmed, needs_validation and rejected. Install is one command: npx skills add https://github.com/cloudflare/security-audit-skill --skill security-audit.

The one number in the README

The README says: "In our test runs, a single run found roughly half of the vulnerabilities that repeated runs found in total." It does not say how many repositories, how many runs or how many vulnerabilities. That is Cloudflare's own result, and the page names no outside party who reproduced it.

Later runs read earlier ledgers and findings to aim at gaps, the README says. The practical reading is that one pass is a sample, and the repo's design assumes several.

Running target code needs a sandbox

The skill will not execute code from the project under audit unless the host enforces limits. The README lists the controls: no external networking, a sanitised allowlisted environment, resource limits, and writes only to assigned scratch paths. Without them, it says, "the workflow keeps the lead as needs_validation instead of executing target code."

An audit run on a laptop with none of those controls therefore produces more unverified leads, not more confirmed bugs. For another tool whose permission settings are not a sandbox, see the site's report on OpenCode's security policy.

Cost is unstated

The README carries no token or dollar estimate. SmartScope, which published a write-up dated October 7 and updated October 13, said the skill is free but that every run launches many sub-agents, so the bill goes to whoever supplies the model. It cited one user's report of about 10% of a weekly allowance on a 20x plan and labelled it unverified.

SmartScope also listed open issues: #64 (ZIP upload failures in the Claude app), #53 (Windows input problems) and #11 (parallel sub-agents stalling in Antigravity). It said no false-positive rate or runtime benchmark has been published.

1. Oct 7 (SmartScope)25K stars
2. Oct 20 (GitHub API)34K stars

Source: SmartScope (Oct 7, 2026) and GitHub API, accessed 2026-10-20

Stars measure interest, not use. The site's earlier report on the agent workspace Orca covers another repo that drew a large star count. The README calls the repo the starting point for a larger harness that "grew into a multi-stage, fleet-wide system", and links a Cloudflare post titled "Build your own vulnerability harness". The site did not fetch that post, so what the internal system finds, and at what cost, is unconfirmed here.

Cloudflare has not said when the repository will next be updated.