Security Audit Skill
cloudflare/security-audit-skillReviewed by AI
What it does
This is an instruction pack for AI coding agents, published by Cloudflare. Once installed, you ask your agent to "security audit this codebase" and it runs a six-step review: it maps the code, sends separate agents to hunt for weaknesses, has a second agent try to disprove each one, then writes a report splitting findings into confirmed, needs checking and rejected. It is for people who already use an AI coding agent; you also need Node installed for the two checker scripts. The repository is only text files and those scripts.
How easy it is to install
C · install 1 or 2 tools · Windows
- Windows C
- macOS C
- Linux C
Same work on all three systems and no installer: you need Node, you run `npx skills add https://github.com/cloudflare/security-audit-skill --skill security-audit` in a terminal, and then ask your AI coding agent for an audit. The README also asks for an OS-enforced sandbox before the agent is allowed to run the audited project's own code, which is extra setup.
Numbers on GitHub
- Stars
- 55,343on Oct 31, 2026
- Downloads
- —all releases
- Growth
- +6,769stars in 5 days
Stars over time
Facts
- Runs on
- Windows, macOS, Linux
- You need to install
- Node.js
- Disk space
- not determined
- Accounts or keys
- None
- Cost
- Free, API costs extra · MIT and free. Nothing runs without an AI coding agent: the README asks for "a coding agent with a model that supports tool use and parallel sub-agents", which you pay for separately, and a long audit uses many model calls.
- License
- MIT (Permissive)
- Can replace
- Snyk
- Last update
- Sep 14, 2026
Risk signals (5)
- Attentionruns the audited project code
README Requirements: "An OS-enforced sandbox for target-controlled builds, tests, processes, browsers, emulators, fuzzers, and fixtures. It must disable external networking, use a sanitized allowlisted environment, enforce resource limits, and allow writes only to assigned scratch paths. Without these controls, the workflow keeps the lead as `needs_validation` instead of executing target code." Setting that sandbox up is on you.
AI review
- Infofinds weaknesses in any codebase you point it at
README Usage: "Start your coding agent in (or pointed at) the codebase you want to audit", and the trigger list includes "pen-test the code". The findings describe real weaknesses, so the report is sensitive material and pointing it at software that is not yours can cross a line.
AI review
- Infomany model calls per run
The README describes six phases with "isolated hunters", "coverage critics", a "fresh verifier" per candidate and "fresh agents" for record verification, and says "a single run found roughly half of the vulnerabilities that repeated runs found in total". That means a lot of AI usage per audit.
AI review
- Infono releases and quiet for weeks
The repository has no GitHub releases and the last commit in this snapshot is 44 days old, so you install whatever is on the main branch at that moment.
AI review
- Infowrites files outside your project
README Usage: "In full audit mode, an unspecified output directory defaults to `~/security-audit-skill/<repo-name>/run-<N>`." Reports with security findings land in your home folder unless you choose another path.
AI review
Evidence is shown in its original language.
What Radar could not check
- Whether the findings are real: the project's own report splits them into confirmed, needs_validation and rejected
- How much your AI agent will charge for a full audit of a real project
- Whether it works on a given coding agent, since it needs parallel sub-agents
- How you would actually set up the OS-enforced sandbox the README requires
- Whether it misses important weaknesses, since the README says one run found about half of what repeated runs found