Cloudflare Security Tool Gains 2,428 Stars in a Day

Cloudflare's security-audit-skill gained 2,428 stars in a single day, according to GitHub's trending page, putting the three-month-old repository at the top of GitHub's technology trending list. The project has drawn more than 18,000 stars in total since Cloudflare published it on June 18.

The skill is a folder of Markdown instructions that turns a coding agent into a security auditor, rather than a full application on its own. Cloudflare said in a blog post announcing the underlying system that it started with a roughly 450-line version of the same skill, run against a single repository, and adjusted the prompts until it surfaced real bugs before building the orchestration around it.

Six phases, from mapping the code to re-verifying the finding

The published skill runs a six-phase audit in one session, according to its GitHub README: three parallel research agents map the architecture and trust boundaries, hunter agents attack the code by class (memory corruption, prompt injection, HTTP framing, tenant isolation), independent validators try to disprove each finding, and a final fresh agent re-verifies survivors against the source before a report is written. Findings are sorted into three verdicts: confirmed, needs_validation, and rejected, the README said.

Installation runs through the Vercel Labs skills CLI: npx skills add https://github.com/cloudflare/security-audit-skill --skill security-audit. The repository is MIT-licensed and needs Node.js to run its two validators.

Cloudflare's own scan generated 20,799 raw candidates

Cloudflare's blog post, describing the fleet-wide system the public skill seeded, gave figures from its internal use across 145 repositories. Of 20,799 raw candidates its Vulnerability Discovery Harness generated, about 12,057 survived validation. Pushed into a shared triage system alongside another harness, the pool reached 13,841 findings. Better context from the recon phase cut the validation rejection rate from 40% to 11%, while the share of high-integrity findings rose from 35% to 58%, Cloudflare said.

1. Before40 %
2. After11 %

Source: Cloudflare blog, accessed 2026-09-21

Cloudflare metric	Value
Raw candidates (VDH)	20,799
Survived validation	12,057
Total in shared triage, 145 repos	13,841
Wishlist requests, 128 repos	25,472

A full scan of a roughly 30,000-line repository takes 3 to 4 hours at that scale, Cloudflare said, though the worst run it measured took just over 14 hours. Cloudflare also said a single pass finds only about half of what multiple passes find, which is why its own pipeline runs the harness repeatedly against the same code.

The independence the instructions require isn't enforced in the public version

French tech outlet Korben reviewed the public skill on Sept. 17 and flagged a gap between what the instructions ask for and what the open-source package can guarantee: "the verifier's independence is required by the instructions, but not enforced." Cloudflare's own deployment enforces that separation with an in-house orchestrator that is not part of the public release, Korben reported.

The skill also refuses to run code it is auditing unless an OS-enforced sandbox is in place, with the network cut off and the target read-only, according to both the README and Korben's account. Without that sandbox, a lead is marked needs_validation rather than followed through. The same problem the skill is built to catch in target code applies to agents more broadly: BragJack showed on Sept. 21 how an AI agent embedded in a browser could be hijacked through the page it was reading, and GitHub's own Copilot runtime rewrite, which ran $120,000 in compute, was itself a bet that agent-adjacent tooling needs sturdier engineering underneath it, not just better prompts.

Cloudflare has not said whether it plans to open-source the orchestrator that enforces verifier independence internally, or how many of the 12,057 validated findings from its own repositories it has disclosed to the maintainers of the affected code.