Security Audit Skill

Tracked

A coding-agent skill from Cloudflare that orchestrates isolated sub-agents through a six-phase security audit pipeline, producing independently verified, machine-readable findings.

Author Cloudflare Open Sourced 2026-06-18 Last Commit Unknown

Overview

Security Audit Skill turns a coding agent into a security auditor. It orchestrates isolated sub-agents through a six-phase workflow — reconnaissance, coverage-led hunting, candidate validation, structured output, independent record verification, and target-neutral reporting — producing findings that are machine-readable and independently verified. It is the single-repo seed of the harness described in Cloudflare's "Build your own vulnerability harness" blog post.

Key Features

  • Six-phase audit pipeline: reconnaissance, coverage-led hunting, candidate validation, structured output, independent record verification, reporting
  • Adversarial validation — the agent that finds an issue never verifies it; a fresh verifier attempts to disprove each candidate
  • Domain-specific hunting playbooks covering memory safety, LLM/prompt injection, web/auth, client-side, supply chain, cloud, RPC, resource exhaustion, data isolation, and desktop/mobile
  • Severity requires demonstrated impact (likelihood × impact), not checklist deviation
  • Additive runs — repeated audits keep finding new vulnerabilities

Use Cases

  • Automating deep security audits of codebases with coding agents
  • Producing machine-readable findings (findings.json) that feed downstream tooling
  • Continuous security review where repeated runs accumulate coverage

Technical Details

  • Recon phase maps architecture, trust boundaries, and input surfaces into architecture.md and a coverage-ledger.json; hunter sub-agents work only from ledger units, with coverage critics identifying gaps
  • Findings written as confirmed / needs_validation / rejected records validated against report-schema.json by zero-dependency Node.js validators (validate-findings.cjs, validate-coverage-ledger.cjs)
  • Sub-agents run in OS-enforced sandboxes (no external networking, sanitized env, resource limits, restricted writes); without a sandbox, target code execution is withheld and leads stay as needs_validation
  • Installed via npx skills add; runs land in ~/security-audit-skill/<repo-name>/run-<N> with REPORT.md, FINDINGS-DETAIL.md, and NEEDS-VALIDATION.md outputs