Cloudflare open-sources a security-audit skill for coding agents

A six-phase audit, verified by agents that never did the hunting, output as JSON. MIT, one command to add. Plus: Edge0 streams a 35B model in under 3GB.

Nowline SEP 22 11:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Your coding agent, now a security auditor

    Cloudflare open-sourced a skill that turns Claude Code, Cursor, or Codex into a vulnerability hunter — install it with `npx skills add https://github.com/cloudflare/security-audit-skill --skill security-audit`, then ask it to audit your codebase. MIT-licensed, and the fastest-climbing repo on GitHub this week.

  • A verifier that never joined the hunt

    The trick that separates it from a 'find bugs' prompt: every candidate vulnerability is handed to a fresh agent whose only job is to disprove it. Findings ship as JSON with three verdicts — confirmed with a full source trace, needs_validation, or rejected — so you chase real bugs, not hallucinated ones.

  • Six phases, machine-readable out

    Recon, coverage-led hunting, candidate validation, structured output, record verification, then a target-neutral markdown report. Zero-dependency Node validators check the JSON automatically, so you can gate on it in code instead of reading a wall of prose.

  • The catch: it wants real compute

    This is not a lint pass. It needs an agent with parallel sub-agents and OS-level sandboxing to run target code safely, plus Node for the validators — and it fans out a lot of sub-agents, so budget for the token bill. Treat it as a discovery engine, not a compliance stamp.

  • Build this weekend: a CI security gate

    Wire the skill into a CI job that runs on every PR and blocks merge when a 'confirmed' finding lands, using its report schema as the contract. Adversarially-verified security review on your own repo, no SaaS scanner subscription.

  • Elsewhere: a 35B model in under 3GB of RAM

    Edge0-35B-A3B streams expert weights off your SSD instead of loading them all, running a 35B mixture-of-experts in ~2.9GB on an M4 Pro at 15-18 tok/s. Apache 2.0 and MLX-backed; Apple Silicon only for now and weak at agentic tool use, but a real 35B on a laptop is the point.