security-audit

The default is "Guidance Mode": when answering security questions or performing partial reviews, it uses only the relevant parts, without creating directories or writing files. It only runs the full six stages when you explicitly request an audit or penetration test of the entire codebase, a comprehensive review, or a report file: reconnaissance of architecture and trust boundaries, dispatching independent agents to find issues based on a coverage ledger, having new agents refute candidates one by one, writing structured records, independent verification, and finally generating a report. SKILL.md stipulates that dependencies are not installed, online services and real accounts are not probed, and only fake identities and fake keys are used.

The default is "Guidance Mode": when answering security questions or performing partial reviews, it uses only the relevant parts, without creating directories or writing files. It only runs the full six stages when you explicitly request an audit or penetration test of the entire codebase, a comprehensive review, or a report file: reconnaissance of architecture and trust boundaries, dispatching independent agents to find issues based on a coverage ledger, having new agents refute candidates one by one, writing structured records, independent verification, and finally generating a report. SKILL.md stipulates that dependencies are not installed, online services and real accounts are not probed, and only fake identities and fake keys are used.

Why it matters Each candidate vulnerability is handed to a new agent to attempt to refute, keeping only those with complete source code evidence; running target code must be done in a sandbox with network isolation and resource limits, otherwise only a verification plan is provided.

skills.sh installs 113867 (+2284 / 24h)

Security Audit: Risk warnings present

  • Initial screening only: Only scanned SKILL.md, no full repository audit performed, installation command not provided for now
  • Audit scope: GitHub cloudflare/security-audit-skill/skills/security-audit/SKILL.md (only scanned SKILL.md text)

Reasoning

  • skills.sh 24 hour installs 2284, cumulative 113867 (represents attention level only)
  • The repository README states this is the initial single-repository version of Cloudflare's vulnerability discovery system, accompanied by an official blog post explanation

Usage

  • After installation, first ask specific security questions (Guidance Mode); for a complete audit, explicitly say "perform a complete security audit on this repository and generate a report," then confirm the scope based on its follow-up questions

Risks

  • The desktop scan only looked at SKILL.md and did not perform a full repository audit; the website does not currently display the installation command; the complete workflow will also read several other stage description files in the repository
  • A complete audit dispatches multiple sub-agents, consuming significant time and tokens; if the local machine lacks a compliant sandbox, it marks issues requiring runtime verification as pending verification

Open · Repository

Best Skills 每日数据(LinklyAI / @BlueeonY,CC BY 4.0)

Explore the latest AI Radar updates