Key takeaways

  • Turn the coding agent you already use into a security auditor that works through a fixed, phased workflow.
  • Fewer false alarms: every candidate goes to a fresh agent whose job is to disprove it.
  • Findings carry one of three verdicts, each with its own evidence rule, so unresolved leads are not passed off as confirmed.
  • Repeat runs add coverage: the project says a single run found roughly half of what repeated runs found.
  • The catches: it needs a capable agent and an OS-enforced sandbox, and the coverage figure is the project's own.

What it is

security-audit is a coding-agent skill that turns your agent into a security auditor. It is released under the MIT licence by Cloudflare, and the code is on GitHub at cloudflare/security-audit-skill.

The problem it solves

Asking an AI agent to look for vulnerabilities tends to give you a pile of unchecked claims. Anyone who has to triage them, from a small team to a security group, pays for the noise.

This skill answers that with structure. It orchestrates isolated agents through reconnaissance, hunting, validation, structured output, independent verification and reporting. The project says it seeded Cloudflare's vulnerability discovery harness, and that it is the single-repo starting point the larger system evolved from.

What you get

  • A defined process: the audit runs in six phases rather than as one open-ended prompt.
  • Independent checks: a fresh verifier tries to disprove each candidate, and the agent that checks a finding is never the agent that found it.
  • Honest verdicts: confirmed, needs_validation and rejected each have their own evidence rule, and a needs_validation record carries no severity.
  • Cumulative coverage: runs against the same repository are additive, using earlier results to target gaps.
  • Less padding: defence-in-depth gaps are not treated as vulnerabilities.

How it compares

The only figure the project offers is its own: in its test runs, a single run found roughly half of the vulnerabilities that repeated runs found in total. That is the maintainers' measurement, not an independent benchmark, and it says nothing about how it fares against other tools.

The catches

  • Agent requirements: it needs a coding agent whose model supports tool use and parallel sub-agents.
  • Sandbox: running target-controlled builds, tests or other code calls for an OS-enforced sandbox. Without one, the workflow keeps the lead as needs_validation instead of executing target code, so results will be less conclusive.
  • Tooling: Node.js is needed for the validators.
  • Output location: the workflow writes inside the target repository only when you select a directory that version control ignores.
  • Evidence: the coverage claim comes from the project's own test runs.

What we did not test

We did not run the skill, point it at any codebase, or check its findings against known vulnerabilities. Every claim here comes from a literal quote in the project's README or the GitHub API, and the descriptions of how it works are the project's own.

Verdict

Editors' view: this is worth a look for developers and security teams who already use coding agents and want audit results that come with evidence and a record of what was checked. If you cannot provide a proper sandbox, expect more unresolved leads. Run it on code you own, and keep a human reviewer in the loop.

Repository
cloudflare/security-audit-skill
JavaScript · MIT
View on GitHub ↗

Sources

Need help adopting this, or have a software project in mind?

Tell us what you are building. The initial consultation is free. Email [email protected] or follow @zolnaltd.

Start a conversation →