Cloudflare Open-Sourced the Skill That Turns Your Agent Into a Security Auditor
Cloudflare put security-audit-skill on GitHub and it went up about 1,250 stars in a day to just under 7,000. It is MIT, it installs with one line, and what it does is run a coding agent through a six-phase security audit instead of asking it nicely to find bugs: reconnaissance, coverage-led hunting, candidate validation, structured output, independent record verification, target-neutral reporting. Repo at https://github.com/cloudflare/security-audit-skill, install with npx skills add, then just say "security audit this codebase."
The design decision worth stealing is that findings get verified by agents that did not discover them. A fresh agent with no stake in the claim re-checks it before it can be marked confirmed, and everything else lands in needs_validation with the open questions written down, or rejected with a record of why it was disproved. That is the fix for the single biggest failure mode of agentic security work, which is not missing bugs, it is a confident stream of plausible findings that eat a human week to triage.
Structured output against a JSON schema is the other half. Reconnaissance writes machine-readable coverage notes, so "did we look at this" becomes a query rather than a vibe, and the hunting phase gets assigned by where coverage is thin instead of wherever the model felt like poking. Anyone who has run an agent at a large codebase knows it will happily audit the same three files forever.
The bigger point is who published it. Cloudflare shipping its internal audit methodology as a skill file, MIT, fourteen commits in, is the strongest evidence yet that the skill format is where security practice gets distributed now. The thing being open-sourced is not code, it is a procedure, and procedures are exactly what agents have been missing. Expect other infrastructure vendors to follow, and expect the first argument about whether a skill that finds a real CVE is a security tool or a security product.
Related reading: [an open model broke into all eleven targets for $4.65](https://clauday.com/article/b6373f31-de17-444b-b500-dd46265457ea)
← Back to all articles
The design decision worth stealing is that findings get verified by agents that did not discover them. A fresh agent with no stake in the claim re-checks it before it can be marked confirmed, and everything else lands in needs_validation with the open questions written down, or rejected with a record of why it was disproved. That is the fix for the single biggest failure mode of agentic security work, which is not missing bugs, it is a confident stream of plausible findings that eat a human week to triage.
Structured output against a JSON schema is the other half. Reconnaissance writes machine-readable coverage notes, so "did we look at this" becomes a query rather than a vibe, and the hunting phase gets assigned by where coverage is thin instead of wherever the model felt like poking. Anyone who has run an agent at a large codebase knows it will happily audit the same three files forever.
The bigger point is who published it. Cloudflare shipping its internal audit methodology as a skill file, MIT, fourteen commits in, is the strongest evidence yet that the skill format is where security practice gets distributed now. The thing being open-sourced is not code, it is a procedure, and procedures are exactly what agents have been missing. Expect other infrastructure vendors to follow, and expect the first argument about whether a skill that finds a real CVE is a security tool or a security product.
Related reading: [an open model broke into all eleven targets for $4.65](https://clauday.com/article/b6373f31-de17-444b-b500-dd46265457ea)
Comments