Audited 2026-10-20 · Sources: GitHub repo metadata, README, SKILL.md + phase docs, file tree. Tabletop audit — nothing downloaded or executed.
The card
| Repo | cloudflare/security-audit-skill |
| Stars | 34.3k (GitHub, audit date) — born October 2026, 14 commits, viral inside a week |
| License | MIT |
| Provenance | Official Cloudflare org — the vendor itself; the README ties it to Cloudflare's "Build your own vulnerability harness" engineering post |
| What it is | A coding-agent skill (SKILL.md + phase docs + two .cjs validators) that runs a six-phase security audit: reconnaissance, coverage-led hunting, candidate validation, structured output, independent record verification, target-neutral reporting |
| Install surface | One skill directory copied into your agent's skills folder. The two validate-*.cjs scripts run under node during a full audit — read them first like any script you execute |
What problem it solves
Agent-run security audits today are vibes: the model wanders the repo, declares a few findings, and nobody can check the coverage. This skill turns the audit into a ledger — deterministic coverage units, isolated hunter agents, a separate verifier whose only job is to disprove each candidate, and machine-readable records (findings.json validated against report-schema.json) that a second pass can independently re-verify. Repeat runs read prior ledgers and hunt the gaps instead of re-reporting the same findings.
Our gate findings
| Gate | Finding |
|---|---|
| Upstream identity | Verified — the actual cloudflare org, MIT, security policy and code of conduct in-tree. This is the single-repo seed of the harness behind Cloudflare's published vuln-hunting pipeline, not a lookalike. |
| License | MIT — clean. |
| Execution discipline | The standout finding. The skill's own rules are tighter than most security products: source inspection is read-only; target code runs only inside an OS-enforced sandbox (no external network, allowlisted environment, read-only target, CPU/memory/disk limits); dummy principals and fixtures only; it explicitly forbids probing deployed endpoints or spending paid API quota. Write-isolation is done with race-safe descriptor promotion — no-follow walks, fstat identity checks, byte limits. Most enterprise audit tooling doesn't go this far. |
| Mode gating | Loads as guidance by default — the full six-phase workflow only fires on an explicit audit/pen-test request, and it asks before creating files when intent is ambiguous. Good containment design. |
| Blast radius | It's instructions, not a daemon — no network calls of its own, no install-time hooks. The real cost surface is compute: the design deliberately spawns fleets of hunter and verifier sub-agents. A full audit on a large repo is a real token bill. Budget before you point it at a monorepo. |
| Findings quality | Verdicts are honest by construction — confirmed needs a full source trace, needs_validation carries no severity until resolved. But the ceiling is the agent running it: a weak model executing this skill produces weak audits, just more organized ones. |
| Maturity | 14 commits — young. Expect file renames and schema drift on pull; pin to a commit if you operationalize it. |
Verdict: GO. Strongest possible provenance (the vendor's own org), MIT, and a security posture strict enough that reading the SKILL.md teaches you sandbox discipline even if you never run it. The only honest costs: it burns tokens by design, and its findings are only as good as the agent holding it.
Best for
Teams already running coding agents who want their security reviews to produce checkable records instead of confident prose — and anyone building an audit pipeline who wants a maintained reference for how the phases should fit.
Also new on the trending list today
- DietrichGebert/ponytail (169.8k★, MIT) — a "laziest senior dev" prompt pack shipped to ~15 agent formats at once (
.claude-plugin,.codex-plugin,.devin-plugin, Cursor, Windsurf, Grok, Kimi and more). No code executes — the blast radius is behavioral: it rewrites how your agent decides what not to build. One repo, fifteen entry points. - Graphify-Labs/graphify (130.3k★, Apache-2.0 + MIT) — deterministic local AST parsing that turns a codebase, docs, SQL schemas and PDFs into a queryable knowledge graph, no vector store. Real engineering: tests, pre-commit, ARCHITECTURE.md, v8. Disclosure: our own fleet already runs it —
graphify-outartifacts exist in the verifiedantiagingclinics repo.
Per-repo deep audits on these two follow on the daily cadence.