Cloudflare Open-Sources a 450-Line Audit Skill, Not the Harness It Became

Cloudflare has published the security-audit skill that seeded its internal vulnerability-hunting system, under an MIT licence, but not the system itself. The repository, cloudflare/security-audit-skill, showed 43.2k stars on its GitHub page on Oct. 24.

The skill is a set of instructions that a coding agent loads when asked for a security review. According to the repository README, it runs six phases, from reconnaissance through independent verification of the final records. A fresh agent tries to disprove each candidate before it is allowed to be called a finding.

What is public and what is not

Grant Bourzikas wrote in Cloudflare's June 18 post, "Build your own vulnerability harness", that the project began as a roughly 450-line skill run against one repository. Turning it into a scanner for 128 repositories took about six weeks. The post says "Hopefully the harness itself will follow shortly."

Piece	Status on Oct. 24	Where it is described
Single-repo skill	Public, MIT licence	GitHub repository
Fleet-wide harness	Not public	Cloudflare blog post
Model names and production prompts	Not disclosed	Cloudflare blog post

The post does not name the models Cloudflare used. It says discovery and validation run on different models so each checks the other's output.

The numbers are Cloudflare's own

The harness produced 20,799 raw candidates, and about 12,057 survived validation, according to the same post. Cloudflare says the initial validation rejection rate fell from 40 percent to 11 percent over time. These are the company's figures from its own research environment, and nobody outside Cloudflare has reproduced them.

1. Raw candidates21K findings
2. Survived validation (about)12K findings

Source: Cloudflare blog, Build your own vulnerability harness, accessed 2026-10-24

The post gives no false-positive rate. The nearest figure is the rejection rate, and it says nothing about findings that passed validation and were still wrong. It also claims no false-negative rate.

What running it costs you

The README sets a hard condition: any target-controlled code must run in an operating-system sandbox with external networking disabled, a sanitized environment and enforced resource limits. Without one, the skill keeps the lead as needs_validation and does not execute it.

Install is one line: npx skills add https://github.com/cloudflare/security-audit-skill --skill security-audit. The agent needs a model that supports tool use and parallel sub-agents.

One run is not enough. The README says that in Cloudflare's test runs a single run found roughly half the vulnerabilities that repeated runs found in total. Fireup.pro, an independent developer site, wrote on Oct. 20 that a full audit launches many sub-agents and uses a lot of model quota, and that model-written fixes in Cloudflare's tests sometimes patched a bug while breaking code that depended on it.

That sandbox requirement is the same problem OpenAI described when it moved frontier training offline after VM escapes, and the one startups such as Daytona now sell a product against.

Stars are not adoption

The star count has moved fast. Fireup.pro reported the repository had passed 33,000 stars on Oct. 20, and the GitHub page showed 43.2k four days later, with 2.7k forks, 32 open issues and 53 pull requests. Stars measure interest, not audits completed.

The README also lists a limit worth reading before a first run: "Defense-in-depth gaps are not vulnerabilities." The skill is built to confirm boundary failures, so a codebase with weak hardening and no clean exploit path will come back looking tidier than it is.

Cloudflare's security-ai-research@cloudflare.com address is the contact listed for the repository. The company has not given a date for the harness release.