Cloudflare Open-Sources AI-Powered Security Audit Framework
Structured six-phase vulnerability detection system for codebases with independent verification
1Automated security auditing framework
Cloudflare's open-source security-audit-skill transforms coding agents into structured vulnerability hunters. The system orchestrates isolated agents through six phases: reconnaissance, coverage-led hunting, candidate validation, structured output, independent verification, and neutral reporting. Originally seeding Cloudflare's internal vulnerability discovery harness, it evolved into a fleet-wide system. The framework maintains rigorous validation checkpoints, running validate-coverage-ledger.cjs after ledger creation/updates and validate-findings.cjs during Phase 4 and post-replacements.
2Three-tiered verdict system
The audit skill classifies findings into distinct verdict categories: 'confirmed' requires complete source tracing and bounded observed results, 'needs_validation' marks exact unresolved facts without severity assessment, and 'rejected' records disproved candidates. This classification ensures only properly validated vulnerabilities progress through the pipeline. Multiple audit runs against the same repository are additive, using prior ledgers to target coverage gaps while carrying forward current-source evidence without treating stale work as covered.
3Comprehensive attack surface coverage
The framework includes specialized hunting classes for diverse vulnerability categories: memory-safety/binary targets, AI/LLM-specific risks (prompt injection, agent tools), web protocols (HTTP framing, authentication), client-side threats (DOM injection, prototype pollution), and cloud deployment risks (IAM, infrastructure-as-code). Each category contains tailored prompts and validation rules. The system also covers supply chain vulnerabilities, RPC/messaging protocols, resource exhaustion scenarios, data isolation issues, and desktop/mobile IPC channels through dedicated hunting classes.
4Validation through adversarial principles
The framework enforces adversarial validation where the agent verifying a finding must differ from the one that discovered it. This prevents confirmation bias in vulnerability detection. Findings require demonstrated boundary failures for confirmation, while potential issues remain as 'needs_validation' with exact unresolved facts. The system emphasizes impact-based severity assessment rather than checklist deviations, recognizing that defense-in-depth gaps constitute hardening notes rather than vulnerabilities.
5Implementation requirements
Running the security audit skill requires a coding agent with tool-use and parallel sub-agent capabilities, Node.js for zero-dependency validators, and OS-level sandboxing for target-controlled processes. The sandbox must enforce network restrictions, sanitized environments, resource limits, and write-only scratch paths. Without proper sandboxing, the workflow maintains findings as 'needs_validation' rather than executing potentially dangerous target code. Output defaults to ~/security-audit-skill/repo-name/run-N unless specified.
6Operational methodology
The skill activates automatically for security-related requests (audit, pen-test, vulnerability detection) operating in either full audit or guidance mode. Full audit mode produces comprehensive report artifacts while guidance mode handles focused security questions. The system maintains strict source-grounding principles, only confirming established boundary failures while tracking unresolved leads precisely. Multiple runs improve coverage significantly - test runs showed single passes detect roughly half the vulnerabilities found through repeated executions.
7Structured output and verification
The framework generates machine-readable findings.json output conforming to report-schema.json, validated by zero-dependency Node.js scripts. Phases 3-6 handle candidate validation, structured output generation, independent record verification, and final reporting. The validate-findings.cjs script performs rigorous checks during Phase 4 and after each Phase 5 replacement, while validate-coverage-ledger.cjs validates ledger integrity throughout the audit process.
8Enterprise-grade security automation
Cloudflare's security-audit-skill represents a significant advancement in AI-assisted code review, combining structured methodology with flexible automation. The multi-phase approach ensures comprehensive coverage while independent verification maintains rigor. The system's ability to learn from multiple runs and its specialized hunting classes for modern attack surfaces make it particularly valuable for complex codebases. As open-source, it provides organizations with enterprise-grade security automation previously limited to internal tools.