Cloudflare open-sources security-audit-skill, a six-phase coding-agent security auditor that seeded its vulnerability harness
- Cloudflare published security-audit-skill, a coding-agent skill that runs a structured audit in six phases: reconnaissance, coverage-led hunting, candidate validation, structured output, independent record verification, and target-neutral reporting.
- The repo has 9.3k stars and 501 forks across 14 commits, and Cloudflare says this skill is the single-repo starting point that grew into its fleet-wide vulnerability discovery harness.
- Output lands in findings.json with three distinct verdicts (confirmed, needs_validation, rejected) validated against report-schema.json, and the parent agent runs validate-coverage-ledger.cjs after creating the ledger and after every later update.
- Repeat runs against the same repo are additive: the skill uses prior ledgers and findings to target gaps, revalidate changed source, and carry forward current-source evidence without treating stale or unresolved work as covered.
- The skill's requirements call for an OS-enforced sandbox, and coverage critics are used inside Phase 2 to find gaps in the hunters' checks.
Hacker News opinions
How many tokens does this eat? I ran it on a fairly small FastAPI project and burned at least 150k before I hit my session limit. Waiting a few hours to keep going.
I threw 1M tokens at it on a medium codebase and got nothing back.
What do you call medium, like 50kloc counting docs?
Cloudflare people, if you're reading this: please consolidate your skills. There are way too many of them on the platform and they pollute our context window.
I'm sure they read HN, someone will see it.
Why not an app store style skills library where each skill gets approved and scanned for security issues first?
14 skills with short descriptions really doesn't seem like a problem to me.
Any clue why an OS-enforced sandbox is in the requirements? Probably so they don't get blamed when the agent does something unexpected and wrecks a machine. Though I doubt normal subscription models will even work with this skill.
Shameless plug, but if this is too many tokens for you, we wrote up the recipe for our own in-house audit skill so it can be replicated and tuned to different environments.
This is awesome, thanks for sharing it.
Tip for security folks using LLMs: audit skills that explicitly frame the task as security research sometimes trigger refusals from the top OpenAI and Anthropic models because they guard against misuse. What works for me is separate skills per bug class with no security framing, plus another skill that combines findings to spot security bugs.
If you're a security professional, just go through their validation and the refusals mostly stop. You still get the occasional downgrade, but not the whole 'I can't write exploits for you' break.