Cloudflare security-audit skillPartly tested🌏 World

Cloudflare's open-source security-audit skill keeps the hunter agent and the verifier agent separate, so the AI that wrote the code does not approve its own work

Asking the same agent session that just shipped a feature to “check security once more” invites self-confirmation. Cloudflare's open-source security-audit skill (GitHub cloudflare/security-audit-skill, MIT) packages their vulnerability-discovery orchestration: six phases (recon → coverage-led hunting → candidate validation → structured output → independent record verification → target-neutral reporting), with hunter/verifier isolation and machine-checked findings.json / coverage-ledger.json via zero-dependency validators. Install with npx skills add https://github.com/cloudflare/security-audit-skill --skill security-audit (optional --global). Chinese-community post @divid_lu49394 (9 Nov 2026) summarized usage and limits (one run covers only part of the surface, static source only, NEEDS-VALIDATION must be reproduced by a human). We cloned main @ c1c8a8c and ran the shipped tests under skills/security-audit/: validate-findings.test.cjs 34/34 pass, validate-coverage-ledger.test.cjs 31/31 pass; SKILL.md confirms guidance mode by default and full six-phase runs only when the user explicitly asks for a full audit. We did not run a live Claude/Codex audit against an app repo (expensive and sandbox-sensitive).

Visit site Cloudflare security-audit skill ↗
Pricing
Free
Region
Global
Where it's from
Overseas
Added
Nov 10, 2026
Tested on
Nov 10, 2026
Version we used
cloudflare/security-audit-skill @ c1c8a8c; Node validator tests
Categories
CodingProductivity
Original author
Cloudflare ↗
Our test
Partly tested
Cost / time
The skill and validators are MIT/free; no Cloudflare account. A full six-phase run spawns many sub-agents and burns substantial Claude/Codex quota; guidance-mode questions are much cheaper.
Source
GitHub cloudflare/security-audit-skill ↗Cloudflare Blog ↗X @divid_lu49394 ↗

The task

Run a structured pre-ship security audit so the same agent is not both author and approver, and keep a machine-validated findings ledger.

Steps

  1. Install with the Skills CLI: npx skills add https://github.com/cloudflare/security-audit-skill --skill security-audit; add --global for a user-level install. Works with Claude Code, Codex and compatible agents; no Cloudflare account required.
  2. Open the target repo and explicitly ask to “security audit this codebase.” Only an explicit full-audit / report request starts all six phases and writes an output directory; ordinary security questions stay in guidance mode.
  3. Full mode writes architecture.md, coverage-ledger.json, findings.json, REPORT.md and related files. The parent re-runs validate-coverage-ledger.cjs after ledger updates and validate-findings.cjs after structured output and record verification.
  4. Read verdicts carefully: confirmed (complete source trace + bounded observed result), needs_validation (missing fact, no severity—human reproduce in a test env), rejected (disproved). Never announce needs_validation as a confirmed vuln.
  5. Re-run important projects; runs are additive against prior ledgers. The skill only sees static source you provide, not live config/runtime. Target code execution requires an OS sandbox (no egress, scratch-only writes, …); if you cannot enforce that, mark needs_validation instead of forcing a run.

Prompt

Install Cloudflare's security-audit skill (https://github.com/cloudflare/security-audit-skill --skill security-audit). In guidance mode first: what inputs and sandbox controls does a full audit of this repo need? Do not create an output directory or start the six-phase workflow until I explicitly confirm.

What the author says

Cloudflare README/blog: the skill is the single-repo starting point of their vulnerability harness, stressing trust boundaries, source-first work, hunter/verifier isolation, plus JSON Schema validators for findings/coverage. @divid_lu49394 (9 Nov 2026): install as above; asking for a security audit triggers the six phases; one run finds roughly half of potential issues, static source only, and NEEDS-VALIDATION items need dual-account reproduction in a test env. The repo showed about 68k stars when we queried it.

Our test

On 10 Nov 2026 we shallow-cloned cloudflare/security-audit-skill @ c1c8a8c (“Clarify guidance and full audit modes”, 14 Sep 2026). Ran node validate-findings.test.cjs and node validate-coverage-ledger.test.cjs under skills/security-audit/. Read README.md and the SKILL.md mode/sandbox sections. Did not npx skills add into a local agent, and did not run the six-phase workflow on an app repo.

What we found when we tried it:validate-findings.test.cjs: 34 tests, 34 pass, 0 fail. validate-coverage-ledger.test.cjs: 31 tests, 31 pass, 0 fail. The skill directory ships SKILL.md, attack-class Markdown files, report-schema.json and both validators; LICENSE is Cloudflare MIT. GitHub API reported stargazers_count 68047 when queried.

How to get in

Public GitHub, MIT, no Cloudflare account. Chinese summary: @divid_lu49394; methodology: Cloudflare blog “Build your own vulnerability harness.” You still need Claude Code / Codex (etc.) quota.

Who it suits

Solo/small teams who want a ledgered security pass on agent-written code before ship. Skip a forced full run if you have no agent quota or cannot offer a local sandbox.

Pitfalls

  • Default mode is guidance, not an automatic full audit—vague “just dig” prompts waste quota or create unwanted report dirs.
  • confirmed / needs_validation / rejected mean different things; do not circulate unverified items as CVEs.
  • Static source only—live config, gateways and real data paths still need humans or other tests.
  • Different from this site's existing iFixAi agent-authz audit: this skill hunts general code vulns, not an agent permission matrix.
  • We only ran the shipped validator tests; we did not execute a live six-phase audit on an app.

More in this category