vibe-check skill
Security checks for apps an agent wrote: rules for AGENTS.md, a live-site scanner and an audit across 17 vulnerability types
- 152 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- The scanner sends about 70 requests to the given address: acceptable only for your own or authorized sites.
- The agent edits code on its own after the audit, so changes must be reviewed in git.
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add skills/benavlabs-vibe-checkDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The scanner sends about 70 requests to the given address: acceptable only for your own or authorized sites.
- The agent edits code on its own after the audit, so changes must be reviewed in git.
This is third-party code. Review the repository files before installing.
What it does
vibe-check protects AI-written projects with four parts. The `AGENTS.md` rules file goes into the project root, where Cursor, Codex, Windsurf or Gemini CLI read it while generating code (Claude Code reads the same text from `CLAUDE.md`). The `skills/vibe-check` skill runs the `check.py` script in Claude Code and then audits the code against a list of 17 vulnerability types. The `check.py` script probes a running site from outside: exposed `.env` and `.git`, public source maps, security headers and CSP, CORS, cookie flags, debug endpoints. It sends about 70 GET, HEAD and OPTIONS requests, prints PASS, FAIL or WARN and gives a fix for each problem. For the audit in `AI-CHECKLIST.md` the agent walks through every category, from Row Level Security and unprotected API routes to IDOR, SSRF, SQL injection and hallucinated packages. It writes a report and a fix plan for each one, applies the fixes and verifies them in the `security/` folder. Manual checks live in `manual-checklist.md`.
Who it is for. Founders and developers who shipped an app built with Cursor, Claude Code or Codex and are unsure whether it leaks keys or has access holes.
Good fit when
- An app is about to go to production and needs a quick outside look at headers, CORS and cookies.
- You suspect an exposed `.env` or `.git` on the server, or source maps left in the build.
- A new project starts with the `AGENTS.md` rules so the agent avoids creating vulnerabilities.
- A full 17-category code audit with a fix plan is required before release.
Not a fit when
- The site belongs to someone else and there is no permission to test it.
- The app already holds real user data and calls for a genuine penetration test.
- You want dependency and git-history secret scanning without your own tools: the workflow calls `gitleaks` and `npm audit` only if they are installed.
Example request
Run a vibe-check security audit on this project. Live URL: https://staging.example.com, the API is at https://api.example.com/health.Limitations
The set does not make an app bulletproof: it covers the basics that took down real projects. Once an app has real users and data, a pentester is needed. The scanner checks only what is visible from outside, and a localhost address gives a partial picture: no HSTS but development source maps. Scan only your own or authorized sites. The agent applies the audit fixes itself, so review code changes in git before committing.
How to disable. Delete the `vibe-check` folder from `.claude/skills` or `~/.claude/skills` and remove `AGENTS.md` or `CLAUDE.md` if you copied the rules. Delete the `security/` folder with audit reports separately.
FAQ
How do I run the scanner without Claude Code?
Run `python3 skills/vibe-check/scripts/check.py https://yourapp.com`. Only Python 3.8 or newer is needed. Add `--api` for an API on another domain and `--json` for machine-readable output.
What requests does the scanner send?
About 70 GET, HEAD and OPTIONS requests to the site, writing no data. Scan only your own apps or ones you are allowed to test.
How does AGENTS.md differ from the audit?
The rules file works while code is written and stops the agent from creating holes. The audit looks for existing problems in the code and on the live site.
Does the set work with Cursor and Codex?
Yes, the rules are copied into `AGENTS.md`. The skill with the automatic scanner targets Claude Code, other agents get the `AI-CHECKLIST.md` file.
Which vulnerabilities does the audit cover?
Seventeen types: from an unprotected database without Row Level Security and open API routes to CSRF, XSS, SQL injection, unverified Stripe webhooks and hallucinated packages. The first five are rated critical.
Discussion
No questions or reviews yet.
Related
SkillsSkill that scans a project for common vibe-coding holes: keys in code, .env in git, open LLM endpoints and weak auth.
A static scanner for vulnerabilities in AI agent code and MCP configs with 66 rules mapped to OWASP Agentic Top 10
A skill for Claude Code and Codex that finds common vulnerabilities in AI-written code and shows a fix for each
VibeSec
VibeSec-Skill
A secure coding skill: the agent looks at code from a bug hunter's perspective and prevents common web vulnerabilities
A local MCP server and CLI find vulnerabilities in AI-generated code and drive a scan, fix and rescan loop, no account needed
A skill that audits vibe-coded apps: secrets, Supabase RLS, authorization, payments, rate limits, mobile and AI integrations
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Repository stars | Main difference |
|---|---|---|---|
| vibe-check (this card) | no | 152 | |
| OWASP Security | no | 440 | checks code against OWASP lists; vibe-check adds a live-site scanner and rules for the agent while it writes code |
| Claude Code Security Reviewer | no | 6,866 | reviews pull request changes through a GitHub Action; vibe-check works on a finished project and a running site |
README in short
The author builds on research into vulnerabilities in AI-written apps and condenses them into 17 types. The first five, from an open database without Row Level Security to secrets in frontend code, are rated critical because they already took down real projects.