Cloudflare Security Audit Skill: Install & Use Guide (2026)
Short answer: security-audit is an open-source agent skill from Cloudflare that turns your coding agent into a security auditor for a codebase. It coordinates isolated sub-agents through six phases. They map the code, hunt for vulnerabilities, try to disprove every candidate, and then write verified, machine-readable findings and reports. By default it only gives security guidance. The full six-phase audit runs only when you ask for one explicitly. It runs target code only inside an OS-enforced sandbox.
Unofficial guide. Antigravity Skills is not affiliated with Cloudflare. The official source is github.com/cloudflare/security-audit-skill. Facts below were checked against the repository on November 11, 2026.
Want it ready to install? Get the Security Audit Skill, a leaner single-agent edition based on Cloudflare's skill, with every safety rule kept and no Node.js dependency. Each download uses 1 credit, and new accounts get 1 free credit.
At a glance
| What it is | An agent skill: a skills/security-audit folder with SKILL.md, 14 reference files, a JSON report schema and two Node.js validators |
| Owner | Cloudflare, Inc. |
| License | MIT: use, change and share it, keeping the copyright and license notice |
| Works with | Any coding agent whose model supports tool use and parallel sub-agents (the skill calls itself agent-neutral) |
| Needs | Node.js for the validators, and an OS-enforced sandbox before any target code is run |
| GitHub stars | 69,149 (November 11, 2026) |
What it does
The README describes a six-phase audit:
- Reconnaissance. Maps architecture, trust boundaries, input surfaces and prior evidence into
architecture.mdandcoverage-ledger.json. - Coverage-led hunting. Assigns isolated hunters to units from the ledger and uses coverage critics to find gaps.
- Candidate validation. Hands every unique candidate to a fresh verifier that tries to disprove it.
- Structured output. Writes
confirmed,needs_validationandrejectedrecords tofindings.jsonand checks them againstreport-schema.json. - Independent record verification. Fresh agents re-check the final source claims.
- Target-neutral reporting. Produces
REPORT.md,FINDINGS-DETAIL.mdandNEEDS-VALIDATION.mdfrom the verified records.
The agent that checks a finding is never the agent that found it. Severity is based on likelihood and impact, not on how far the code strays from a checklist. Runs against the same repository add up: later runs use earlier ledgers and findings to target gaps.
Cloudflare says this skill seeded its internal vulnerability discovery harness, described in Build your own vulnerability harness.
How to install it
The README documents one install method, the Skills CLI. It does not give separate instructions for each agent. Run npx skills --help for agent-selection and non-interactive options.
npx skills add https://github.com/cloudflare/security-audit-skill \
--skill security-auditFor a user-level installation, add --global:
npx skills add https://github.com/cloudflare/security-audit-skill \
--skill security-audit \
--globalIf you prefer to copy the folder by hand, the skill is the skills/security-audit folder of the repository. Cloudflare does not document manual installation. The folder locations below are the standard skill paths for each agent.
Google Antigravity
Cloudflare does not list Antigravity by name. To install by hand, copy skills/security-audit into ~/.gemini/config/skills/ (global) or .agents/skills/ inside a project. The older ~/.gemini/antigravity/skills/ path is still read by the IDE. The full audit needs parallel sub-agents, so check that your setup can delegate tasks before you start a full audit.
Claude Code
Cloudflare does not list Claude Code by name. To install by hand, copy skills/security-audit into ~/.claude/skills/.
Codex
Cloudflare does not list Codex by name. To install by hand, copy skills/security-audit into ~/.agents/skills/.
More detail on skill folders for each agent: How to install skills.
How to use it
Open your agent in the codebase you want to audit and ask in plain words. These examples come from the README:
security audit this codebase
find security vulnerabilities in ./src
do a security review, output to ~/audits/my-projectThe skill has two modes:
- Guidance mode (the default). Security questions, focused reviews and work on specific findings use only the relevant parts of the skill. It does not create an output directory or write audit files unless you ask for report artifacts.
- Full audit mode (explicit request only). Runs all six phases and writes the report files. It starts only when you explicitly ask to audit or pen-test a codebase, ask for a full, comprehensive or end-to-end review, or ask for report artifacts. If your request could mean either mode, the skill asks you one question first.
Unless you name a directory, full audit output goes to ~/security-audit-skill/<repo-name>/run-<N>, outside your repository. It writes inside the target repository only when you pick a directory that version control ignores.
Things to know before you run it
- You need a sandbox to execute code. Reading source is always read-only. Builds, tests, browsers, fuzzers and fixtures from the target may only run inside an OS-enforced sandbox. That sandbox must block external networking, start from an allowlisted environment, enforce CPU, memory and time limits, and allow writes only to assigned scratch paths.
- Without a sandbox, fewer findings are confirmed. If those controls are not available, the skill does not run target code. Leads that need a run to prove them stay as
needs_validation, together with a safe plan for checking them. - It stays local. The skill tells the agent not to probe deployed endpoints, external services, production identities or other users' data. It also tells the agent not to install dependencies or let builds fetch them.
- One run is not enough. In Cloudflare's test runs, one pass caught only about half of the issues that several passes caught together. Run it more than once.
- Hardening notes are not vulnerabilities. If one layer already blocks an attack, a missing second layer is recorded as a hardening note.
- Full audits use parallel sub-agents. Check that your agent and model support them before you start a full audit.
- The install command runs
npx. It downloads the Skills CLI from npm. Review the repository before installing.
FAQ
Is Cloudflare's security-audit skill free?
Yes. It is open source under the MIT license. Running it uses your own agent and model, which may cost you money.
Will it start a full audit as soon as it is installed?
No. The skill defaults to guidance mode. The six-phase workflow and its report files run only when you explicitly ask for an audit, a pen test, a full review or report artifacts.
Do I need a sandbox?
You need one only if you want the skill to execute target code to confirm findings. Without an OS-enforced sandbox it keeps to reading source and marks unproven leads as needs_validation.
Where do the reports go?
By default, to ~/security-audit-skill/<repo-name>/run-<N>. You can name a different output directory in your request.
Does it work in Google Antigravity?
Cloudflare describes the skill as agent-neutral but does not name Antigravity. It needs an agent that supports tool use and parallel sub-agents.