Run Cloudflare's security-audit skill

Cloudflare's security-audit skill is a free MIT skill that drives a coding agent through a source audit. A full run spends agent invocations, and the skill file tells the agent to stop when the budget cannot cover them.

Share link · Share on X

Last updated Oct 14, 2026

mediumConfiguration

Shows up in: Next.js

Markdown version

What it is

Cloudflare's security-audit skill is a set of instructions for a coding agent, not a website you paste a URL into. The README says it turns the agent into an auditor that maps the codebase, hunts for failures of a real trust boundary, and writes findings only after a different agent tries to disprove them. The README also says this skill is the single-repo starting point for the vulnerability harness described in Cloudflare's post Build your own vulnerability harness.

The skill file splits work into guidance mode and full audit mode. Loading the skill does not start the full workflow. A full audit runs when you explicitly ask for an audit, a pen-test, or report files. The six phases in the README are reconnaissance, coverage-led hunting, candidate validation, structured output, an independent check of those records, and a report built from the verified records.

How to check in 5 minutes

Open the README and confirm the install command is npx skills add https://github.com/cloudflare/security-audit-skill --skill security-audit.
Run npx skills --help and look for --agent. On 2026-10-14 the examples included --agent claude-code. The skill file itself says the skill is agent-neutral and does not name Claude Code.
Read the "Cost budget" section of skills/security-audit/SKILL.md before you ask for a full audit. It counts spend as agent invocations, not as a dollar price.

How to fix

Claude Code

Install into Claude Code with the agent flag shown in the Skills CLI help, checked on 2026-10-14:

npx skills add https://github.com/cloudflare/security-audit-skill \
  --skill security-audit \
  --agent claude-code

Start Claude Code in the repository you want reviewed. The README's examples are "security audit this codebase" and "find security vulnerabilities in ./src". Ask for guidance mode if you only want a question answered. Ask for a full audit only when you mean the six phases.

Before a full audit, set a budget. The skill file says one coverage unit is about one hunter assignment, and a surviving candidate is one or two verifier assignments. It also says the agent must not launch hunters if the remaining budget cannot reserve the critic and verifier calls, and that a strict budget must not be exceeded silently. The README does not name a credit unit or a dollar price. The cost is whatever your coding agent charges for those invocations.

Next.js

The skill reads source. It does not, by itself, open your deployed URL. The skill file says not to probe deployed endpoints, and to leave a lead as needs_validation when the decisive fact is outside the repository. A missing header on the live site can stay unconfirmed until you check the response yourself.

Read findings.json by the verdict names in the README. confirmed has a source trace and an observed result. needs_validation has an unresolved fact and no severity. rejected was disproved. Do not treat needs_validation as a vulnerability, and do not merge a change the agent describes. The skill file says the audit describes fixes and does not modify the target.

Copy-paste prompt

Paste this into Cursor, Claude Code, or whichever agent is editing the repo. Read the diff before you merge it.

You are using Cloudflare's security-audit skill in guidance mode. Do not start a full six-phase audit unless I set an agent-invocation budget first.

1. Read SKILL.md and tell me whether my request is guidance mode or full audit mode.
2. If it is a full audit, estimate hunter and verifier invocations from the cost-budget section and stop if I have not named a maximum.
3. Do not probe the deployed site. If a check needs a live response, mark it needs_validation.
4. Report confirmed, needs_validation, and rejected separately. Do not give needs_validation a severity.

Questions

Does the README say this spends Claude credits?

No. The README does not mention credits or a price. SKILL.md tells the parent agent to record a budget as a maximum number of agent invocations, and to refuse the run when that budget cannot pay for reconnaissance plus the reserved critic and verifier calls.

Is one run enough?

The README says that in Cloudflare's test runs, a single run found roughly half of the vulnerabilities that repeated runs found in total. Treat one report as partial. The skill file says a quick or scoped run must present itself as partial coverage.