Cloudflare Security Audit Skill - Setup Guide¶

Source: cloudflare/security-audit-skill via skills.sh (25.6K installs, first seen Jun 18, 2026; 26,403⭐ on GitHub, MIT LICENSE; last pushed Sep 14, 2026) Skill: security-audit (single-skill repo; originally from 89jobrien/steve, adapted and maintained by Cloudflare) Category: Security / Code Auditing Quality Tier: 🟢 Production (official Cloudflare publisher; battle-tested origin as the seed of Cloudflare's vulnerability discovery harness; MIT LICENSE; active development; Gen Agent Trust Hub Pass / Socket Pass / Snyk Warn - verified Oct 8, 2026)

The official Cloudflare skill that turns any coding agent into a security auditor. It orchestrates isolated agents through reconnaissance, coverage-led hunting, candidate validation, structured output, independent record verification, and target-neutral reporting. It seeded Cloudflare's vulnerability discovery harness, described in Build your own vulnerability harness: the harness grew into a multi-stage, fleet-wide system, and this skill is the single-repo starting point it evolved from. The skill is agent-neutral (works with any agent platform) and MIT-licensed; per skills.sh it was originally from 89jobrien/steve before Cloudflare adapted it.

Installation¶

# Single skill install (this repo ships one skill)
npx skills add cloudflare/security-audit-skill --skill security-audit

# GitHub URL form, as shown in the skill README
npx skills add https://github.com/cloudflare/security-audit-skill --skill security-audit

# User-level installation
npx skills add cloudflare/security-audit-skill --skill security-audit --global

This is a single skill, not a suite: there is no wildcard family install, just security-audit. Requirements: a coding agent with tool use and parallel sub-agents, Node.js for the zero-dependency validators, and (for full audits) an OS-enforced sandbox for target-controlled builds, tests, processes, browsers, emulators, fuzzers, and fixtures.

The Six-Phase Workflow¶

Phase	Output	What Happens
1. Reconnaissance	`architecture.md`, `coverage-ledger.json`	Map architecture, trust boundaries, input surfaces, prior evidence, and deterministic coverage
2. Coverage-led hunting	hunter records plus critic gap reports	Assign isolated hunters from ledger units, record their checks, and use coverage critics to find gaps
3. Candidate validation	validated candidates	Give every unique candidate to a fresh verifier that tries to disprove it
4. Structured output	`findings.json` validated against `report-schema.json`	Write `confirmed`, `needs_validation`, and `rejected` records
5. Independent record verification	verified records	Fresh agents verify final source claims; material replacements receive another independent verifier
6. Target-neutral reporting	`REPORT.md`, `FINDINGS-DETAIL.md`, `NEEDS-VALIDATION.md`	Derive reports from the verified records and the coverage ledger

Helper scripts (zero-dependency, run with Node.js): validate-coverage-ledger.cjs runs after the ledger is created and after each later ledger update; validate-findings.cjs runs in Phase 4 and again after every Phase 5 replacement.

Verdict Semantics¶

Verdict	Meaning
`confirmed`	Complete source trace plus a bounded observed result
`needs_validation`	Exact unresolved fact; no severity attached
`rejected`	Candidate disproved during verification

Multiple runs against the same repo are additive: prior ledgers and findings target gaps, changed source gets revalidated, and stale or unresolved work is never treated as covered.

Operating Modes¶

Mode	When It Applies	Behavior
Guidance (default)	Security questions and focused vulnerability work	Uses only the relevant parts of the skill; no automatic full workflow or artifact creation
Full audit	An explicit request to audit or pen-test a codebase	Runs the six-phase workflow and writes report artifacts; an unspecified output directory defaults to `~/security-audit-skill/<repo-name>/run-<N>`

Full audit mode writes inside the target repository only when you explicitly select a directory that version control ignores.

Key files: SKILL.md (setup, core principles, anti-patterns) plus hunting-class references such as RECONNAISSANCE.md, HUNTING.md, ATTACK-CLASSES.md, MEMORY-SAFETY-AND-BINARY.md, AI-AND-LLM.md (prompt-injection and agent/tool hunting), WEB-PROTOCOL-AND-AUTH.md, CLIENT-SIDE.md, SUPPLY-CHAIN-AND-RELEASE.md, CLOUD-AND-DEPLOYMENT.md, and VALIDATION-AND-REPORTING.md.

Why This Matters for Hermes Agents¶

Agents shipping real code need a structured, evidence-first security review loop, not ad hoc pattern matching. This skill is the public distillation of a production vulnerability discovery harness that grew fleet-wide at Cloudflare, packaging the expensive lessons (isolated hunters, adversarial verification, auditable coverage) into a workflow any agent can run on any codebase. It is agent-neutral and MIT-licensed, so it drops into a Hermes fleet with no platform lock-in.

Usage¶

Start the agent in (or pointed at) the codebase you want audited, then trigger the skill:

security audit this codebase
find security vulnerabilities in ./src
do a security review, output to ~/audits/my-project

The skill activates automatically on matching requests (security audit, find vulnerabilities, pen-test the code, security review). Focused security questions stay in guidance mode; a direct audit or pen-test request switches to full audit mode. Re-run against the same repo to grow coverage - Cloudflare notes that in its test runs a single run found roughly half of the vulnerabilities that repeated runs found in total.

Verification¶

# Confirm the skill is installed locally
hermes skills list | grep security-audit

# Confirm the validators are present in the skill directory
ls validate-coverage-ledger.cjs validate-findings.cjs

Smoke test: point the agent at a small repo and ask "security audit this codebase". In full audit mode it should create coverage-ledger.json during reconnaissance and later produce findings.json, REPORT.md, FINDINGS-DETAIL.md, and NEEDS-VALIDATION.md in the output directory. Cross-check install counts and per-skill verdicts on the skills.sh page linked above.

Security¶

skills.sh security verdicts for security-audit (verified Oct 8, 2026) - check the skill's security page on skills.sh before production use:

Check	Verdict
Gen Agent Trust Hub	Pass
Socket	Pass
Snyk	Warn

Note: the skill writes files, and in full audit mode may execute target-controlled code paths (builds, tests, processes, browsers, emulators, fuzzers, fixtures). Run it on a branch or worktree, keep output out of version control, and use an OS-enforced sandbox before executing target code.

Integration with CorpusIQ¶

CorpusIQ is the connector layer that lets AI assistants read the systems a business already runs on - accounting, payments, CRM, email, files and analytics - and answer with the source named. It is read-only by design: it retrieves data and never writes to, or acts in, a connected system, which is what makes it safe to point at live accounts.

This page is part of the CorpusIQ directory, the reference we maintain for the MCP server and agent-skill ecosystem. Run alongside CorpusIQ, a third-party server like this one can be combined with a company's own business data in a single answer.

See what CorpusIQ connects

Limitations¶

Two modes matter: guidance mode (the default) does not run the full workflow or create artifacts. Full audit mode must be explicitly requested.
needs_validation is not a severity level. It records an exact unresolved fact for follow-up, so it should not be triaged as a reported vulnerability.
Full audit mode writes report artifacts and may execute target code paths. Plan for a branch or worktree, an ignored or external output directory, and the README's sandbox requirements (no external networking, sanitized allowlisted environment, resource limits, scratch-only writes).
Runtime is heavyweight by design: a required coding agent with parallel sub-agents plus Node.js for the validators. This is an audit harness, not a quick linter.
Snapshot data, verified Oct 8, 2026: 25.6K skills.sh installs; 26,403 GitHub stars; MIT LICENSE; last pushed Sep 14, 2026. Counts will drift over time.
Authorship note: per the skills.sh listing, the skill was originated by 89jobrien as 89jobrien/steve and adapted by Cloudflare; this catalog entry documents the Cloudflare-maintained version.

Pro Tips¶

Run a guidance-mode focused pass on your highest-risk area (auth, input parsing, agent tool handling) before committing to a full six-phase audit - it is fast and often surfaces where the real risk lives.
Keep audit output out of version control: choose an ignored directory or a path outside the repo, and run on a branch or worktree since artifacts land on disk.
Re-run against the same repo instead of treating one pass as complete. Runs are additive: prior ledgers target gaps and revalidate changed source.
Do not skip the validators: validate-coverage-ledger.cjs after each ledger update and validate-findings.cjs after Phase 5 replacements keep artifacts schema-clean.
Treat needs_validation entries as a follow-up queue of exact unresolved facts; clearing them on the next run is where repeated audits earn their keep.

← Skills Catalog | Skills Marketplace → Powered by CorpusIQ