Cloudflare Security Audit Skill
Turns your coding agent into a security auditor with a multi-phase pipeline for finding exploitable vulnerabilities.
Evidence: The skill is designed as read-only audit, not modifying code, but permission boundaries are not explicitly declared; user confirmation for output directory is required, but not for high-risk actions; data flow is clearly described (input codebase, output reports), but no mention of external uploads; sensitive data handling not specified; dependencies minimal (Node.js only), but no dependency list or vulnerability scan; external effects limited to file generation, but write scope not explicit; no rollback mechanism; source attribution via README and LICENSE, publisher unverified. Deductions: unclear permission boundaries, insufficient user confirmation, no sensitive data handling, no rollback.
Evidence: Documentation is internally consistent, phases clearly described; dependency only Node.js, but version unspecified, availability moderate; failure messages not detailed, only validator mentioned. Deductions: insufficient failure messages.
Evidence: Targeted at security audit scenarios, audience clear; capability boundaries clear (only report exploitable vulnerabilities); trigger words explicit; environment requirements clear (coding agent with tool use and parallel sub-agents). Deductions: none significant.
Evidence: Clear file structure, good information architecture; detailed installation notes; stable naming; examples and FAQ via README; known limitations mentioned (single run covers half); MIT license; no versioning or changelog; maintenance responsibility via contact email. Deductions: no versioning/changelog.
Evidence: Output is structured JSON and reports, high usability; high marginal value (automated security audit); cost-benefit reasonable (zero dependencies, only Node.js). Deductions: none significant.
Evidence: Claims based on blog and testing, but no specific evidence provided; no cross-source verification; fact-inference separation good (validation phase). Deductions: claims lack traceability, no cross-source verification.
- The skill generates files and may overwrite existing ones; ensure running in an isolated environment.
- The skill does not explicitly state whether codebase content is sent to external services; review its data flow.
- Publisher identity is unverified; use with caution.
What does this agent do, and when should you use it?
The security-audit-skill is a coding-agent skill from Cloudflare that automates security audits. It orchestrates multiple parallel agents through six phases: recon, hunting, validation, reporting, structured output, and independent verification. Each phase has dedicated prompt files (e.g., RECONNAISSANCE.md, HUNTING.md) and supports many attack classes, including web, client-side, AI/LLM, and memory-safety. Outputs include human-readable reports and a findings.json conforming to a JSON schema, validated by a zero-dependency Node.js script. The skill is the basis for Cloudflare's vulnerability harness, emphasizing actual exploitability and adversarial validation. It requires a coding agent with tool use and parallel sub-agent support.
The skill guides a coding agent through a six-phase audit via prompt files (SKILL.md, RECONNAISSANCE.md, etc.). In recon, parallel agents map architecture. In hunting, parallel agents attack the code from angles like injection, access control, and business logic, spawning sub-agents. Validation uses separate agents to disprove findings. Reporting produces REPORT.md and FINDINGS-DETAIL.md. Structured output writes findings.json per report-schema.json, validated by validate-findings.cjs. Independent verification uses fresh agents to verify factual claims. Multiple runs are additive, reading prior findings to skip known issues.
- A security engineer auditing a large codebase for exploitable vulnerabilities before release.
- A development team integrating automated security scans into their CI pipeline, requiring machine-readable findings.
- A researcher conducting in-depth audits of open-source projects, seeking structured and verifiable results.
- An LLM application developer checking for prompt injection or agent-tool vulnerabilities.
- A contractor performing security reviews for multiple clients, needing consistent methodology.
What are this agent's strengths and limitations?
- Adversarial validation uses independent agents to eliminate false positives.
- Systematic six-phase pipeline covers diverse attack classes, including AI/LLM-specific ones.
- Structured findings.json conforms to JSON Schema, enabling machine parsing and integration.
- Multiple runs accumulate to increase vulnerability coverage.
- Requires an advanced coding agent with tool use and parallel sub-agents, limiting model choices.
- Requires Node.js for the validation script.
- Documentation mandates concrete attack scenarios for every finding, increasing manual review overhead.
- As a skill, it depends on the host agent platform and lacks standalone execution.
How do you install or deploy this agent?
Install via Skills CLI:
npx skills add https://github.com/cloudflare/security-audit-skill --skill security-auditAdd --global for a user-level install. Node.js is required for the validator.
How do you use this agent?
Start your coding agent in the codebase to audit and issue a command like "security audit this codebase". The skill triggers automatically and asks for an output directory (default ~/security-audit-skill/<repo-name>/run-<N>). Ensure your agent model supports tool use and parallel sub-agents.