AGENTCONN

// Skill profile

Cloudflare Security Audit Skill

Claude Code Advanced Security Open Source

cloudflare/security-audit-skill is an open-source coding-agent skill for automated multi-phase security audits. Derived from Cloudflare's internal vulnerability discovery harness, it orchestrates a fleet of isolated sub-agents through six phases: reconnaissance (mapping architecture and trust boundaries), coverage-led hunting (assigning isolated hunters to ledger units with coverage critics), candidate validation (fresh verifier disproves each candidate), structured output (confirmed/needs_validation/rejected records to findings.json), record verification (fresh agents re-verify source claims), and reporting (REPORT.md, FINDINGS-DETAIL.md, NEEDS-VALIDATION.md). Each finding is machine-readable and independently verified by a separate agent pass, reducing false positives. Includes attack-class playbooks covering memory safety, LLM/agent systems, web protocols and auth, supply chain, cloud/IaC, and more. Additive multi-run support reuses prior ledgers. +2,979 stars/day on GitHub Trending. 40K stars, 2.5K forks. MIT licensed.

Input / Output

Accepts

codebase repository-url

Produces

security-report json-findings markdown-report

Overview

Cloudflare Security Audit Skill is the open-sourced starting point of Cloudflare’s internal vulnerability discovery harness. Instead of running a single security scan, it orchestrates multiple isolated sub-agents through a six-phase pipeline: reconnaissance (mapping the attack surface), hunting (searching for vulnerability classes), validation (confirming exploitability), reporting (structuring findings), structured output (machine-readable JSON), and independent verification (a separate agent pass that re-checks each finding).

The multi-phase approach with independent verification is the key differentiator. Most automated security tools produce long lists of potential issues with high false-positive rates. By having a separate verification agent independently confirm each finding, the skill produces reports where flagged issues are more likely to be real, exploitable vulnerabilities rather than theoretical concerns.

Key Capabilities

Six-phase audit pipeline: The skill breaks security auditing into six distinct phases: (1) Reconnaissance maps architecture, trust boundaries, and input surfaces into architecture.md and coverage-ledger.json. (2) Coverage-led hunting assigns isolated hunters to ledger units with coverage critics to find gaps. (3) Candidate validation uses a fresh verifier to disprove each unique candidate. (4) Structured output writes confirmed, needs_validation, and rejected records to findings.json, checked against a JSON schema. (5) Record verification re-verifies source claims with fresh agents. (6) Reporting generates REPORT.md, FINDINGS-DETAIL.md, and NEEDS-VALIDATION.md.

Adversarial verification: The verifier is never the finder. Every finding goes through a separate verification pass by an agent that did not participate in the initial discovery. confirmed requires a complete source trace and bounded observed result. needs_validation covers leads blocked by an exact unresolved fact.

Attack-class playbooks: Built-in coverage for memory safety, LLM/agent systems, web protocols and auth, client-side code, supply chain and releases, cloud and IaC, RPC and messaging, resource exhaustion, data isolation, and desktop/mobile IPC.

Additive multi-run support: Later runs reuse prior ledgers and findings, revalidate changed source, and do not treat stale work as covered. The authors report a single run finds roughly half the vulnerabilities that repeated runs find in total.

Machine-readable output: Findings are structured as JSON with severity ratings, affected code locations, reproduction steps, and remediation suggestions. Zero-dependency Node.js validators for the findings file and coverage ledger.

Agent-agnostic installation: Installs via the Skills CLI (npx skills add). Works with any coding agent whose model supports tool use and parallel sub-agents.

Use Cases

Security-conscious development teams use the skill to run automated security audits before releases. Open-source maintainers use it to audit contributions and dependencies. Penetration testers use it as a starting point for more targeted manual testing. DevSecOps teams integrate it into CI pipelines for continuous security validation.

Considerations

The skill requires a capable underlying model to perform effective security analysis — it orchestrates the audit but the quality of findings depends on the model’s security knowledge. Running a full six-phase audit on a large codebase consumes significant tokens due to the multi-agent architecture. The independent verification phase adds cost but substantially improves signal-to-noise ratio. An OS-enforced sandbox is required for running target code (builds, tests, browsers, fuzzers) — without it, leads stay as needs_validation and target code is not executed. MIT licensed, no Cloudflare account required.

Who It’s For

Security engineers who want automated, structured vulnerability discovery. Development teams that need security audits integrated into their agent workflows. Anyone building on Cloudflare’s reference harness pattern for multi-agent security analysis.

Tags

#security #audit #vulnerability-discovery #pentesting #code-review #multi-agent #cloudflare #adversarial-validation #attack-playbooks

Compatible Agents

AI agents that work well with Cloudflare Security Audit Skill.

Similar Skills