Cloudflare has released an open-source coding-agent skill designed to organize security reviews into a repeatable, evidence-driven workflow. The project, called Security Audit Skill, is presented as the single-repository starting point from which Cloudflare developed a broader vulnerability-discovery harness.
The skill divides an audit into six phases covering reconnaissance, coverage-led investigation, candidate validation, structured reporting, independent record checks and neutral presentation of results. Rather than treating a model's initial suspicion as a vulnerability, the workflow records what code was examined and requires later validation before a finding can be marked confirmed.
Machine-readable artifacts are central to the design. A parent agent runs a coverage-ledger validator after the ledger is created and again whenever later work changes it. A separate findings validator runs during the fourth phase and after replacements made in the fifth. Those repeated checks are intended to keep the audit record internally consistent as multiple agents investigate different parts of a codebase.
The reporting model distinguishes three outcomes. A confirmed finding needs a complete trail to the relevant source and a bounded observed result. A needs-validation record must identify the precise unresolved fact and cannot assign a severity. A rejected record preserves a candidate that investigation disproved. Keeping rejected and unresolved work visible can reduce repeated effort and makes the boundary between evidence and inference clearer.
Audits are additive across runs. The skill can read earlier coverage ledgers and findings, use them to locate gaps, and recheck source that has changed. It can also carry forward evidence tied to code that remains current, while avoiding the assumption that stale or unresolved work has been fully covered. This approach is meant to support repeated reviews rather than a single disposable scan.
The project offers two operating modes. A direct request to audit or penetration-test a codebase invokes the full process, while narrower security questions and focused vulnerability work use a guidance mode unless the user asks for report artifacts. In a full audit with no selected destination, output goes to a numbered directory outside the target repository. Writing audit material inside the repository requires an explicitly chosen directory that version control ignores.
Cloudflare describes the release as a foundation for agent-assisted review, not an automatic guarantee that a codebase is secure. Its safeguards focus on traceability: agents may generate hypotheses, but validators, source-linked records and explicit verdict states control what enters the final report. That separation is particularly important in security work, where an unsupported high-severity label can be as misleading as a missed defect.
