Cloudflare security-audit-skill: Turn a Coding Agent Into a Security Auditor

Cloudflare's official security-audit-skill runs a six-phase audit with isolated hunters, adversarial verification and machine-readable findings. The verbatim install, the three verdicts, the sandbox it needs, and how to run it alongside DeepSeek Harness.

Last updated: 2026-11-17

Quick answer

security-audit is Cloudflare's own coding-agent skill for running a structured security audit — not a scanner you point at a URL. The repository summary is: "A coding-agent skill for multi-phase security audits with independently verified, machine-readable findings." It is official: the repo sits in the cloudflare GitHub organization, is MIT licensed, and its README links the Cloudflare engineering post Build your own vulnerability harness, describing this skill as the thing that seeded that harness.

At the GitHub API snapshot of 2026-11-17 the repository had 74,073 stars and 4,962 forks, created 2026-06-18 and last pushed 2026-09-14. Read the README at that date and the version you install may differ; it is a young repository that moved between June and September 2026.

The part that separates it from "ask the model to look for bugs" is the design rule in its own README: the agent that checks a finding is never the agent that found it.

The six phases

Each phase writes files you can inspect afterwards, which is why the output survives the session.

PhaseWhat happensFiles
1. ReconnaissanceMap architecture, trust boundaries, input surfaces, prior evidence and deterministic coveragearchitecture.md, coverage-ledger.json
2. Coverage-led huntingAssign isolated hunters from ledger units, record their checks, use coverage critics to find gapsledger updated in place
3. Candidate validationEvery unique candidate goes to a fresh verifier that tries to disprove it—
4. Structured outputWrite verdicts and validate them against the schemafindings.json
5. Independent record verificationFresh agents verify final source claims; material replacements get another verifierfindings.json updated
6. Target-neutral reportingDerive the reports from verified records and the coverage ledgerREPORT.md, FINDINGS-DETAIL.md, NEEDS-VALIDATION.md

The validators are part of the workflow, not optional linting. The README says the parent runs validate-coverage-ledger.cjs after creating the ledger and after each later update, and validate-findings.cjs in Phase 4 and again after every Phase 5 replacement. Both are "zero-dependency", which is why the requirements list only needs Node.js.

The three verdicts

The distinctions are the useful part, because they map onto how security work actually stalls:

  • confirmed — a complete source trace and a bounded observed result.
  • needs_validation — an exact unresolved fact, and no severity.
  • rejected — a disproved candidate, recorded so nobody re-litigates it.

The README's design principles explain the discipline behind that: only established boundary failures get confirmed; severity requires impact (likelihood × impact, not "deviation from a checklist"); defence-in-depth gaps are not vulnerabilities — if Layer A prevents the attack, the absence of Layer B is a hardening note.

Install it

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

For a user-level install:

npx skills add https://github.com/cloudflare/security-audit-skill \
  --skill security-audit \
  --global

The README points at skills.sh for agent-selection and non-interactive flags, and npx skills --help lists them locally.

Run it

Start your coding agent in — or pointed at — the codebase you want audited, then ask in plain language. The README's own examples:

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

The skill activates when the request matches its trigger (security audit, find vulnerabilities, pen-test the code, and similar). A direct audit or pen-test request uses full audit mode; security questions and focused vulnerability work use guidance mode unless you ask for report artifacts.

In full audit mode, an unspecified output directory defaults to ~/security-audit-skill/<repo-name>/run-<N>. The workflow only writes inside the target repository when you explicitly choose a directory that version control ignores — worth remembering on a repo you care about.

What it needs

  • A coding agent whose model supports tool use and parallel sub-agents — the phases depend on spawning isolated hunters and fresh verifiers.
  • Node.js, for the two validators.
  • An OS-enforced sandbox for anything target-controlled: builds, tests, processes, browsers, emulators, fuzzers, fixtures. The README's bar is specific: disable external networking, use a sanitized allowlisted environment, enforce resource limits, allow writes only to assigned scratch paths.

That third requirement is the one people skip. Without it the workflow degrades by design — it keeps the lead as needs_validation rather than executing code from the target. The sandbox and approval model in dsh is covered on sandboxes, approvals and secrets; read it before pointing this at a third-party repo.

Running it alongside DeepSeek Harness

The skill is a folder containing a SKILL.md plus its phase documents — RECONNAISSANCE.md, HUNTING.md, VALIDATION-AND-REPORTING.md, and attack-class files such as AI-AND-LLM.md, CLIENT-SIDE.md and CLOUD-AND-DEPLOYMENT.md. That is the same object dsh loads from .dsh/skills/, so the install is a copy:

git clone https://github.com/cloudflare/security-audit-skill
mkdir -p .dsh/skills && cp -r security-audit-skill/skills/security-audit .dsh/skills/

Two practical notes for a dsh run:

  1. Budget the parallelism. The phases want isolated sub-agents. dsh's subagents guide covers how to run them and what they cost in tokens.
  2. Keep the default output directory. Leaving reports out of the repo (~/security-audit-skill/...) means the audit cannot pollute a diff or leak a finding into a commit. Review findings like you would a pull request — code review is the right habit to bring.

The repository also ships validate-findings.test.cjs and validate-coverage-ledger.test.cjs, so the validators themselves are tested; that is a small signal about how seriously the project takes machine-readable output.

What it is not

  • Not a replacement for a pentest or a bug bounty. It is an agent workflow that produces reviewable records. Whether a finding is exploitable in production still depends on deployment context the model cannot see.
  • Not a one-shot guarantee. The README states that multiple runs are additive, and that in Cloudflare's own test runs "a single run found roughly half of the vulnerabilities that repeated runs found in total". That is the project's self-reported result on its own runs — not an independent benchmark, and it says nothing about your codebase.
  • Not a substitute for the sandbox. The requirements section is a precondition, not a suggestion. Read safety first.
  • Not a licence to skip review. Every verdict is designed to be checked by a human; needs_validation exists precisely so unresolved facts do not get laundered into severity.

FAQ

Is cloudflare/security-audit-skill an official Cloudflare project? Yes — it lives in the cloudflare GitHub organization, is MIT licensed, and its README links Cloudflare's engineering blog post about the vulnerability harness it seeded.

How do I install it? npx skills add https://github.com/cloudflare/security-audit-skill --skill security-audit, with --global for a user-level install.

What does a run produce? findings.json with confirmed / needs_validation / rejected records validated against report-schema.json, plus REPORT.md, FINDINGS-DETAIL.md and NEEDS-VALIDATION.md, all derived from architecture.md and coverage-ledger.json.

Do I need a sandbox? Yes for anything target-controlled: no external networking, allowlisted environment, resource limits, writes only to scratch paths. Without it the workflow keeps leads as needs_validation instead of executing target code.

Can I run it from DeepSeek Harness? Yes — copy the security-audit skill folder into .dsh/skills/. It needs tool use plus parallel sub-agents, and Node.js for the validators.

Sources

Every claim above comes from these pages, read on 2026-11-17:

  • github.com/cloudflare/security-audit-skill — README: the repository summary, the six phases and their output files, the three verdicts and their definitions, the two zero-dependency validators and when they run, the install commands and the --global variant, the usage examples, full-audit vs guidance mode, the default output directory and the version-control-ignored condition, the requirements list including the OS-enforced sandbox, the design principles, the "roughly half of the vulnerabilities" self-reported result, and the contact address. GitHub API snapshot taken 2026-11-17: MIT, Organization owner cloudflare, 74,073 stars, 4,962 forks, created 2026-06-18, pushed 2026-09-14 (a snapshot, not a fixed number).
  • blog.cloudflare.com/build-your-own-vulnerability-harness — linked from the README as the write-up of the harness this skill seeded.
  • skills.sh — the Skills CLI the README names for installation and agent selection.

The skill is a Cloudflare open-source project; it is not a DeepSeek product and is not affiliated with this site. Commands come from the README at the date shown — this repository changed between June and September 2026, so re-check before running.

Related reading: sandboxes, approvals and secrets, code review with dsh, installing skills, subagents.

DSH Plugins is an independent community directory of DeepSeek Harness plugins. Not affiliated with or endorsed by DeepSeek. Third-party plugins are not security-audited — review the source before installing.

New DeepSeek Harness plugins, weekly. No spam.