<- all tokdocs

Cloudflare's Security Audit Skill: What It Does and What It Found

By , founder of Mavric and Apso

View on TikTok ->

security-audit-skill is a free, MIT-licensed coding-agent skill from Cloudflare that runs a multi-phase security audit on a codebase and has a separate agent try to disprove every finding. A TikTok from @ai.honeycove presents it as a brand-new tool that found over 7,000 bugs at Cloudflare. The skill is real and useful, but it was released on 18 June 2026, and the 7,245 findings came from a larger internal system that Cloudflare built on top of it and did not release.

What is Cloudflare's security-audit-skill?

The repository describes itself as "A coding-agent skill for multi-phase security audits with independently verified, machine-readable findings." It is a single skills/security-audit/ directory with a SKILL.md, companion documents for each phase and attack class, a report schema and two small Node.js validators.

It was published alongside the Cloudflare blog post Build your own vulnerability harness by Dan Jones, Alexandra Godoi and Grant Bourzikas, dated 18 June 2026. The video was posted on 12 October 2026, 116 days later, and calls it "just" open sourced.

The "completely free" part is accurate. The repository is MIT licensed. The model inference it drives is not free, and the blog post says "Running hundreds of agents over a fleet of repos is not cheap" without giving a dollar or token figure.

Did this skill find 7,000 bugs at Cloudflare?

Not by itself. The blog post says, "Out of 20,799 raw candidates generated by VDH, only about 12,057 survived validation," and "Ultimately this left 7,245 actionable findings for engineering teams to act on." VDH is the Vulnerability Discovery Harness, which Cloudflare paired with a Vulnerability Validation System and ran across 128 repositories.

The released skill is the seed of that system. The post says, "We started with a ~450-line security-audit skill that we ran on a single repository," and that the release is "the initial skill we used to develop the harness." The README says the same: "The harness grew into a multi-stage, fleet-wide system; this skill is the single-repo starting point it evolved from."

The video's order of events is right, since Cloudflare built it for itself first. The 7,245 number belongs to the fleet-wide harness, not to the single-repository skill you install.

How many GitHub stars does it have?

The video says the project "has already crossed 20,000 stars on GitHub." The GitHub API returned 32,473 stars and 2,028 forks on 2026-10-19. Star-history.com shows a best rank of #1 on 17 September 2026. The claim is true as a floor but understates the count.

How does the adversarial validation work?

This is the most accurate part of the video, which says each bug is "handed to a separate agent whose only job is to disprove that bug exists." The README's Phase 3 reads "give every unique candidate to a fresh verifier that tries to disprove it," and a design principle says "The agent that checks a finding is never the agent that found it." Phase 5 adds another pass where fresh agents verify the final source claims.

The video then says "only the real issues make it to the finish line." The skill actually writes three verdicts to findings.json:

  • confirmed: a complete source trace and a bounded observed result
  • needs_validation: an exact unresolved fact and no severity
  • rejected: a disproved candidate

Unresolved leads do reach the report, in a separate table. The reporting guide says to "Keep these as prioritized leads without severity." The README also notes that "a single run found roughly half of the vulnerabilities that repeated runs found in total," so one pass will miss a lot.

The claim that you get "every verified issue and exactly how to fix it" holds. Each confirmed finding in REPORT.md must include the "smallest source fix," and FINDINGS-DETAIL.md must include "source-level remediation and regression case."

How to install security-audit-skill

The video says to install it into Claude Code. The README is agent-neutral and asks for:

  • "A coding agent with a model that supports tool use and parallel sub-agents"
  • "Node.js for the zero-dependency findings and coverage-ledger validators"
  • An OS-enforced sandbox for running target code

The install command uses the Skills CLI:

npx skills add https://github.com/cloudflare/security-audit-skill --skill security-audit

Add --global for a user-level install.

Why the sandbox requirement matters

The video leaves out the sandbox, and it decides whether you get confirmed findings. The README requires "An OS-enforced sandbox for target-controlled builds, tests, processes, browsers, emulators, fuzzers, and fixtures," with external networking disabled, a sanitized environment, resource limits and writes limited to scratch paths. "Without these controls, the workflow keeps the lead as needs_validation instead of executing target code."

Running it on an unsandboxed laptop will produce leads rather than confirmed bugs. Expect long runs too. The blog post says "A full scan of a complex repo can take hours; the worst run took just over 14 hours."

Is it only for vibe-coded apps?

No. The word "vibe" does not appear in the README, and Cloudflare ran the derived system on 128 of its own production repositories. The video names logins, databases, APIs and prompt injection. The repository also has attack-class documents for memory safety and binaries, supply chain and release signing, cloud and infrastructure-as-code, tenant isolation, and desktop, mobile and local IPC. A memory-safety and kernel hunting class was added on 26 June 2026 by an outside contributor.

The security bug cards and terminal output in the video are designed animations. The video shows no real run of the tool.

Bottom line

Cloudflare's security-audit-skill is a free, single-repository audit skill with real adversarial validation and fix-oriented reports. It was released in June 2026, it is the seed of the system that found 7,245 issues rather than that system itself, and it needs an OS-level sandbox to confirm findings. Plan on multiple runs and paid model usage.

Sources

Published October 12, 2026. Based on a TikTok by @ai.honeycove. Full video notes (markdown)