Cloudflare security-audit-skill: stars vs real feedback
The repository sits at 65,554 stars (as of 10 Nov 2026) and 4,260 forks, and the discussion I can cite is a sliver of that. Six mentions pass the opinion test: two Hacker News comments from one thread on 2026-09-17 and four GitHub issues dated between 2026-09-16 and 2026-11-02. That gap is the first thing to know about cloudflare/security-audit-skill, an MIT-licensed JavaScript skill that tells a coding agent to run a multi-phase security audit and write its findings as machine-readable records. The repo was created on 2026-06-18, and the last push in the collected facts is 2026-09-14, so all six opinions came after the most recent code change on record.
The Hacker News side is about the shape of the project, not about results. @vntok, 2026-09-17 defended keeping the prompts in git, pointing to a subdirectory of two dozen files and around 300Kb of text, and wrote that "Storing/visualizing small text changes over time as revisions is exactly what Git is excellent at". @nicce, 2026-09-17 was doubtful about model access and said "I doubt that OpenAI models with normal subscription, for example, wont even work with this skill." That is a guess about subscriptions, and it does not describe a failed run.
The GitHub issues carry the harder evidence, though all four are still open and issues skew toward trouble. @MubashirullahD, 2026-10-13 opened issue #77 about a run on their own codebase and reported that "After 2.5 hours, the output is hidden by Codex", then asked which model to use. @HetCreep, 2026-09-16 wrote up issue #20, a small blind comparison with one seeded target and three arms: this skill, their own multi-lens review pipeline, and a plain single-agent control. The excerpt in the bundle stops before any result, so I cannot say which arm did best. @miikkij, 2026-10-21 described issue #88, a full audit run on a multi-tenant platform with about 30 coverage units, two hunting waves and 50 confirmed findings, followed by a second step that the excerpt cuts off. @ornemeolans, 2026-11-02 filed issue #130 after a full audit of an app backed by managed Postgres (Supabase), saying the reconnaissance agent reported that the migrations contain no row-level security policies. The rest is truncated.
Set against 117 open issues on the repo, six citable voices is thin, and the sample leans toward people who had run full audits. None of the excerpts settles whether the reported findings held up under review. Next to Semgrep or CodeQL, which run fixed rules or queries and return the same answer twice, this skill hands the hunting to model agents, and its README says repeated runs add to earlier ones, so a clean result and a missed bug can look identical until someone re-checks.
Want a second opinion on security-audit-skill?
Ask an AI that can read this page โ one click and you get its take on security-audit-skill.
