What Is the Cloudflare security-audit Skill? Auditing Your Own Code for Security With a Coding Agent

Instead of asking a coding agent for a single answer to "are there security vulnerabilities in this code?", the security-audit skill ties the job to a method. According to the README's design principles, a finding's severity is based on likelihood × impact, not on deviation from a checklist; a claim that can't be validated doesn't count as a finding. To do this, the skill runs the agent like an audit team and forces every claim through evidence and independent validation. A skill is a ready-made instruction package that tells the assistant how to do a specific job.

Cloudflare wrote this skill to find vulnerabilities in its own codebase; it later grew the internal system that scans 128 repositories out of this skill, and open-sourced the skill with a blog post on June 18, 2026. Who it helps: developers and small technical teams who want the code of their site, app, API or library audited before release. What it doesn't do: it doesn't fix your code or scan your live site or server; it works by reading the source code and only with safe local testing.

!This skill is for defensive use: only audit your own code or code you're authorized to audit.

There are two operating modes. According to SKILL.md, the skill is in guidance mode by default: for a security question, a review of a single file or the investigation of a specific finding, it uses only the relevant part, doesn't run the six phases and doesn't create files. Full audit mode turns on only when you explicitly ask it to audit a codebase, run a "pen-test", do a comprehensive review or produce a report file. If your request looks like both, the agent asks a single question before creating files.

Source: README "What it does", SKILL.md "Operating modes" and "Full audit workflow", RECONNAISSANCE.md (October 29, 2026). The Cloudflare blog describes the first version as seven phases; the seventh phase was submission to Cloudflare's own internal system. The current version in the repository has six phases.

The README gives the installation with the Skills CLI: Vercel's command-line tool that installs skills into the skill folders of different coding agents. No account or API key is needed. You run the command in the folder of the project you'll audit; the CLI asks which agent to install it for.

• For agent selection and non-interactive install options, the README points to the npx skills --help command.
• The Skills CLI requires Node.js 22.20 or later; the skill's validator scripts also run on Node.js.
• There's no version tag in the repository: the command takes the main branch as of install time (on October 29, 2026, the latest change was September 14, 2026, commit c1c8a8c). Later changes come in with npx skills update.

The Skills CLI sends anonymous usage data during installation; to turn it off, put DISABLE_TELEMETRY=1 in front of the command (Skills CLI README).

The official README doesn't describe uploading it to the Claude app (claude.ai, Desktop) as a ZIP; an upload error reported by a user is under "Common problems".

Start your agent inside the code repository you want to audit (or point it to the repository) and ask for a security audit in plain language. According to the README, when the request matches a trigger (such as "security audit", "find vulnerabilities", "pen-test the code"), the skill kicks in automatically; you don't need to type a separate command. The requests below are examples from the README; the outputs are described according to the official documentation.

Running it again pays off. A second run on the same repository reads the previous coverage ledger and findings, and targets gaps and changed code. README: in Cloudflare's own tests, a single run found about half of the vulnerabilities that repeated runs found in total (the maker's measurement).

• REPORT.md: run profile and scope, a short security posture summary, a table of confirmed findings (severity, title, affected boundary, one-line outcome), each finding's source location, local reproduction, impact and smallest source fix, a separate "needs validation" table, hardening notes and a coverage summary.
• FINDINGS-DETAIL.md: the full trace of medium, high and critical findings, local reproduction steps with a fake user and resource, a fix and a suggested regression test.
• NEEDS-VALIDATION.md: the list of claims that couldn't be validated, the fact each one is stuck on and how to resolve it.

A clean run can end with zero findings; in that case the documentation explicitly asks not to invent "low" findings. The skill doesn't change the target code: it describes the fix, and applying it is up to you (or an agent you ask separately).

Source: README, VALIDATION-AND-REPORTING.md (phases 4–6), SKILL.md "Separate priority from certainty" (October 29, 2026).

Cloudflare's documentation names no specific agent; it describes the skill as "agent-agnostic" and sets one requirement: a coding agent with a model that can use tools and run parallel subagents. The README of the install tool, Skills CLI, supports OpenCode, Claude Code, Codex, Cursor and 75 more agents.

!An audit report isn't a security guarantee: the documentation explicitly states that a single run doesn't exhaust the target and that areas left out of scope shouldn't be considered "clean".

The skill itself is free and open source (MIT); no account, subscription or Cloudflare product is needed. The real cost is that the audit uses your agent's model: a full audit runs many subagents (reconnaissance agents, a hunter per coverage unit, one or two validators per candidate). The official sources give no token or dollar amount for an audit; the amount depends on the repository's size, the profile and your agent's plan.

• Track spending in your agent's usage dashboard; do your first try on a small repository with the quick profile.
• Write an agent-call limit into your request; SKILL.md asks that the budget you set not be silently exceeded.
• The Cloudflare blog says that in its own multi-repository system almost all of the compute budget went to the hunting phase; this isn't a figure for the single-repository skill.

Source: LICENSE (MIT, Copyright (c) 2025-2026 Cloudflare, Inc.), SKILL.md "Cost budget", Cloudflare blog "What it costs" (October 29, 2026). This is not legal advice.

The command comes from the command table in the Skills CLI README (npx skills remove [skills...]). The audit outputs stay in a separate folder (by default ~/security-audit-skill/); you delete them yourself.