Cloudflare's Security Audit Skill is a free MIT skill that makes a coding agent audit a codebase in six phases and check each finding with a fresh agent.

What it is: an audit procedure for your coding agent

The repo, cloudflare/security-audit-skill, holds one skill, security-audit. Cloudflare's README says it is the skill that seeded the company's own vulnerability discovery harness, which grew into a multi-stage, fleet-wide system, and that this skill is the single-repo starting point it evolved from. It does not scan by itself: your agent runs the steps, using parallel sub-agents.

Six phases: reconnaissance maps the architecture, trust boundaries and input surfaces into architecture.md and coverage-ledger.json. Coverage-led hunting assigns isolated hunters to ledger units, and coverage critics look for gaps. Every unique candidate then goes to a fresh verifier that tries to disprove it. Results are written to findings.json and checked against report-schema.json. Fresh agents verify the final source claims. Reports REPORT.md, FINDINGS-DETAIL.md and NEEDS-VALIDATION.md come last.
Three verdicts:confirmed needs a complete source trace and a bounded observed result. needs_validation names the exact unresolved fact and carries no severity. rejected records a candidate that was disproved.
Hunting guides by target type: separate files cover memory safety and binaries, AI and LLM targets, HTTP protocol and auth, client-side code, supply chain and release, cloud and deployment, RPC and messaging, resource exhaustion, data isolation, and desktop, mobile and local IPC.
Repeat runs add up: a later run on the same repo uses earlier ledgers and findings to aim at gaps and recheck changed source, without counting stale or unresolved work as covered.

Two modes, and a sandbox rule

Guidance mode is the default. For security questions, focused reviews or triage it uses the relevant parts only, creates no output directory and writes no audit files.
Full audit mode runs when you ask to audit or pen-test a codebase, want a comprehensive review or ask for report files. If a request could mean either, the skill says it asks one question first.
Source inspection is read-only. The skill says to run target-controlled builds, tests, browsers or fuzzers only in an OS-enforced sandbox with no external network, an allowlisted environment, a read-only target and tight CPU, memory and time limits. If any of those cannot be enforced, it does not run the target's code and keeps the lead as needs_validation with a safe plan.
It tells the agent to use dummy accounts and secrets, and not to probe live endpoints, shared infrastructure or other people's data.

Install and use

Install with npx skills add https://github.com/cloudflare/security-audit-skill --skill security-audit, and add --global for a user-level install.
Open your coding agent in the codebase and say, for example, "security audit this codebase", "find security vulnerabilities in ./src" or "do a security review, output to ~/audits/my-project".
Unless you name a folder, full audit output goes to ~/security-audit-skill/<repo-name>/run-<N>, outside the repository. It writes inside the repo only if you pick a folder that version control ignores.
You need an agent whose model uses tools and parallel sub-agents, and Node.js for the two validators that check findings.json and the coverage ledger. The README does not name specific agents; the skill is written agent-neutral.

What Cloudflare says, and what it does not show

The README's principles: only confirm established boundary failures, the agent that checks a finding is never the agent that found it, severity needs real impact, and a missing second layer of defense is a hardening note, not a vulnerability.
Cloudflare says that in its own test runs a single run found about half of the vulnerabilities that repeated runs found in total. That is the company's statement about its tests, not a result Seailife saw, and the README gives no detection rate for any codebase.
The independent checking is a design written into the skill's instructions. Nothing in the README promises that nothing was missed, so treat the output as input for a human reviewer, not a clean bill of health.

How fresh it is

This is not a new repository. It was created on June 18, 2026, and its last push was on September 14, 2026, a commit that clarified the guidance and full audit modes. It has had no update in the weeks since.
On November 7 it had about 64,500 stars, about 4,200 forks and 115 open issues and pull requests together.
The name is unrelated to Cloudflare Clef, which is a decision model on Workers AI, not a security tool.
Seailife did not install the skill or run an audit with it. The phases, modes and install steps above are from the README and SKILL.md.

What it is: an audit procedure for your coding agent

Six phases: reconnaissance maps the architecture, trust boundaries and input surfaces into architecture.md and coverage-ledger.json. Coverage-led hunting assigns isolated hunters to ledger units, and coverage critics look for gaps. Every unique candidate then goes to a fresh verifier that tries to disprove it. Results are written to findings.json and checked against report-schema.json. Fresh agents verify the final source claims. Reports REPORT.md, FINDINGS-DETAIL.md and NEEDS-VALIDATION.md come last.

Three verdicts:confirmed needs a complete source trace and a bounded observed result. needs_validation names the exact unresolved fact and carries no severity. rejected records a candidate that was disproved.

Hunting guides by target type: separate files cover memory safety and binaries, AI and LLM targets, HTTP protocol and auth, client-side code, supply chain and release, cloud and deployment, RPC and messaging, resource exhaustion, data isolation, and desktop, mobile and local IPC.

Repeat runs add up: a later run on the same repo uses earlier ledgers and findings to aim at gaps and recheck changed source, without counting stale or unresolved work as covered.

Two modes, and a sandbox rule

Guidance mode is the default. For security questions, focused reviews or triage it uses the relevant parts only, creates no output directory and writes no audit files.

Full audit mode runs when you ask to audit or pen-test a codebase, want a comprehensive review or ask for report files. If a request could mean either, the skill says it asks one question first.

Source inspection is read-only. The skill says to run target-controlled builds, tests, browsers or fuzzers only in an OS-enforced sandbox with no external network, an allowlisted environment, a read-only target and tight CPU, memory and time limits. If any of those cannot be enforced, it does not run the target's code and keeps the lead as needs_validation with a safe plan.

It tells the agent to use dummy accounts and secrets, and not to probe live endpoints, shared infrastructure or other people's data.

Install and use

Install with npx skills add https://github.com/cloudflare/security-audit-skill --skill security-audit, and add --global for a user-level install.

Open your coding agent in the codebase and say, for example, "security audit this codebase", "find security vulnerabilities in ./src" or "do a security review, output to ~/audits/my-project".

Unless you name a folder, full audit output goes to ~/security-audit-skill/<repo-name>/run-<N>, outside the repository. It writes inside the repo only if you pick a folder that version control ignores.

You need an agent whose model uses tools and parallel sub-agents, and Node.js for the two validators that check findings.json and the coverage ledger. The README does not name specific agents; the skill is written agent-neutral.

What Cloudflare says, and what it does not show

The README's principles: only confirm established boundary failures, the agent that checks a finding is never the agent that found it, severity needs real impact, and a missing second layer of defense is a hardening note, not a vulnerability.

Cloudflare says that in its own test runs a single run found about half of the vulnerabilities that repeated runs found in total. That is the company's statement about its tests, not a result Seailife saw, and the README gives no detection rate for any codebase.

The independent checking is a design written into the skill's instructions. Nothing in the README promises that nothing was missed, so treat the output as input for a human reviewer, not a clean bill of health.

How fresh it is

This is not a new repository. It was created on June 18, 2026, and its last push was on September 14, 2026, a commit that clarified the guidance and full audit modes. It has had no update in the weeks since.

On November 7 it had about 64,500 stars, about 4,200 forks and 115 open issues and pull requests together.

The name is unrelated to Cloudflare Clef, which is a decision model on Workers AI, not a security tool.

Seailife did not install the skill or run an audit with it. The phases, modes and install steps above are from the README and SKILL.md.

Cloudflare Security Audit Skill

What it is: an audit procedure for your coding agent

Two modes, and a sandbox rule

Install and use

What Cloudflare says, and what it does not show

How fresh it is

Comments

Cloudflare Security Audit Skill

What it is: an audit procedure for your coding agent

Two modes, and a sandbox rule

Install and use

What Cloudflare says, and what it does not show

How fresh it is

Comments