security-audit-skill: Multi-stage security audit for coding agents
security-audit-skill is an open-source coding agent skill by Cloudflare that transforms AI coding agents into structured security auditors. It performs deep code repository audits through six phases: reconnaissance, coverage-guided hunting, candidate verification, structured output, independent record verification, and target-neutral reporting. The output is machine-readable and independently verifiable. Key differentiators include multi-agent isolated collaboration, strict verification processes, and cumulative multi-run mechanisms, effectively reducing false positives and ensuring audit traceability. Originating from Cloudflare's internal vulnerability discovery system, it serves as a foundation for large-scale automated security auditing, suitable for continuous, repeatable codebase assessments, especially in CI/CD pipelines or as a base tool for security research.
Background and Context
Traditional code security audits rely on manual effort that struggles to scale across modern attack surfaces, while automated scanners often flood teams with false positives, leading to alert fatigue. Cloudflare’s open-source security-audit-skill addresses this gap by transforming AI coding agents into structured, multi-stage security auditors. Rather than a standalone scanner, it is a skill pack that integrates with coding agents, drawing on Cloudflare’s internal vulnerability discovery system described in the blog “Build your own vulnerability harness.” Released as a single-repository starting point on GitHub, the project has quickly garnered over 18,000 stars, signaling strong community interest in AI-assisted, repeatable security auditing.
The skill targets engineering teams seeking to shift security left and embed continuous auditing into development workflows. It is designed to be loaded into coding agents that support custom skills, with clear setup instructions in SKILL.md. By orchestrating isolated agents through a rigorous six-phase pipeline, security-audit-skill aims to produce machine-readable, independently verifiable findings that reduce hallucination-driven false positives and enable incremental audits over time.
Deep Analysis
The framework’s core is a six-phase audit pipeline. Phase 1, Reconnaissance, maps the target’s architecture, trust boundaries, and input surfaces, generating an architecture.md and a coverage-ledger.json that tracks audit coverage. Phase 2, Coverage-Guided Hunting, dispatches isolated hunting agents that record checked items against the ledger; a coverage review agent then identifies gaps. Phase 3, Candidate Verification, assigns each unique vulnerability candidate to a fresh agent that actively attempts to falsify it, minimizing confirmation bias. Phase 4, Structured Output, writes results to findings.json and validates them against a report-schema.json, classifying each as confirmed (with full source trace and observable result), needs_validation (unresolved facts but no severity), or rejected (falsified). Phase 5, Independent Record Verification, deploys a new agent to verify final source claims; any material substitution triggers another round of independent verification. Finally, Phase 6, Target-Neutral Reporting, derives REPORT.md, FINDINGS-DETAIL.md, and NEEDS-VALIDATION.md from the verified records and coverage ledger.
Validation scripts run at critical junctures: validate-coverage-ledger.cjs after ledger creation and each update, and validate-findings.cjs after Phase 4 and after each Phase 5 substitution. This ensures data consistency. Key differentiators include multi-agent isolation to avoid single-perspective bias, independent verification that curbs hallucinated findings, and a cumulative multi-run mechanism that re-audits only gaps and changed code, avoiding redundant work. The skill also bundles extensive attack-category files covering memory safety, AI/LLM prompt injection, web protocol authentication, client-side DOM injection, supply-chain dependencies, cloud deployment configuration, RPC serialization, and more, reflecting deep audit coverage across diverse technology stacks.
Industry Impact
security-audit-skill lowers the barrier to deep code auditing, granting small and mid-sized teams access to vulnerability-discovery capabilities once reserved for large enterprises. Its design promotes a shift from one-time scans to continuous, repeatable assessments, potentially influencing the architecture of future static analysis (SAST) tools. Integration into CI/CD pipelines is straightforward: each code commit can trigger an incremental audit that focuses only on new or modified attack surfaces, producing structured outputs suitable for defect tracking or compliance evidence.
The project’s rapid community adoption—evidenced by its GitHub star count—and Cloudflare’s active maintenance suggest it could become a foundational tool for AI-driven security research. By making audit processes programmable and extensible, it encourages a standardized approach to software supply-chain security, where findings are not only machine-readable but also independently verifiable, reducing the trust gap between automated tools and human reviewers.
Outlook
Current limitations include dependency on the underlying coding agent’s model quality, which can affect audit depth, and potentially long execution times for large repositories, necessitating careful scoping of audit units. Human review remains essential, especially for business-logic flaws that automated agents may misjudge. Future development may focus on integration with agent frameworks like LangChain or CrewAI, community-contributed attack-category libraries, and automatic linking of findings to vulnerability management platforms. As Cloudflare feeds more experience from its internal harness into the open-source skill, it has the potential to become a cornerstone of AI-powered security auditing, driving the industry toward continuous, verifiable code assessment.