Vibecheck— 为AI构建的应用提供上线前安全审计

main

Branches Tags

Go to file

Code Open more actions menu

Latest commit

History

9 Commits

9 Commits

Folders and files

Name Name Last commit message

Last commit date

.claude-plugin

.claude-plugin

.github

.github

assets

assets

examples

examples

hooks

hooks

skills

skills

test

test

.gitattributes

.gitattributes

.gitignore

.gitignore

CHANGELOG.md

CHANGELOG.md

CODEOFCONDUCT.md

CODEOFCONDUCT.md

CONTRIBUTING.md

CONTRIBUTING.md

LICENSE

LICENSE

README.id.md

README.id.md

README.md

README.md

SECURITY.md

SECURITY.md

package.json

package.json

View all files

Repository files navigation

Your AI wrote the app. vibecheck makes sure it doesn't leak your users' data, or your wallet.

Install · What it catches · Commands · Guard · CI · 🇮🇩 Bahasa Indonesia

In January 2026, a vibe-coded social network exposed 1.5 million API tokens and 35,000 email addresses . One Supabase key sat in the front-end, and Row Level Security was off ( Wiz , Infosecurity ). In 2025, 170+ Lovable apps shipped with the same hole ( CVE-2025-48757 ). Every week, someone finds out their OpenAI key was in the JavaScript bundle by reading their bill.

AI tools write code that works . Nobody checks that it's safe . vibecheck is the senior engineer who reviews your app before you hit deploy, in plain language, with the fixes ready to paste.

📄 See a full sample report

What you get

🚦 A Launch Score and a verdict. 0 to 100, plus 🔴 NO-GO, 🟡 FIX FIRST or 🟢 GO, so you know if you can ship today.

🔍 A real audit, not regex soup. A zero-dependency scanner (60+ checks) finds the leads fast. Your AI agent then reads the code to verify each one, traces attack flows like "can user A read user B's data?", and drops the false positives.

🗣️ Plain language. Every finding starts with what can happen ("anyone can read every private message"), then the technical detail. Reports come in the language you write in.

📋 Copy-paste fix prompts. Every finding has a ready prompt for Claude Code, Cursor, Lovable, Bolt, v0 or Replit, so it helps even if your app lives in a no-code builder.

🛠️ /vibecheck:fix . Fixes issues one at a time, criticals first, verifies each one, and never "fixes" by weakening security.

🛡️ An always-on guard. A hook stops your agent from hardcoding API keys or committing .env before it happens.

🚀 /vibecheck:deploy . Catches "works on my machine, broken in production" before you deploy: env vars missing on the host, hardcoded localhost URLs, imports whose letter case only works on a Mac, packages missing from package.json, servers on the wrong port, SQLite on serverless, and the dashboard steps people forget.

🧭 /vibecheck:unstuck . Breaks the "AI fix loop" when the same bug keeps coming back.

What it catches

Area Examples

🔑 Secrets OpenAI / Anthropic / Stripe / AWS / Supabase keys in code; secrets behind NEXTPUBLIC / VITE / EXPOPUBLIC_ ; .env committed now or in git history

🗄️ Database Supabase tables without RLS, using (true) policies, users editing their own isadmin , policies trusting usermetadata ; Firebase "test mode" and open rules; the service-role key in browser code

🔐 Auth API routes and server actions without a server-side check, IDOR (reading other people's records), admin by client flag, middleware-only auth, jwt.decode without verify

💸 Money AI calls from the browser ( dangerouslyAllowBrowser ), paid APIs with no rate limit ("denial of wallet"), prices taken from the request, unverified Stripe / Midtrans / Xendit webhooks, premium unlocked from a success URL

🧱 Input SQL injection, mass assignment, XSS via dangerouslySetInnerHTML and AI output, SSRF, shell injection, prompt injection through tools

🚀 Config CORS * with credentials, debug mode, public source maps, disabled TLS checks, missing security headers, hallucinated npm packages

🛟 Reliability No git, no backups, no error monitoring, schema not in migrations, no lockfile

🌍 Launch Privacy policy and terms, account deletion (App Store), OG tags, sitemap, robots, 404, favicon, analytics, email DNS

🚀 Deploy Env vars not set on the host, localhost URLs, import './button' vs Button.tsx (breaks on Linux), undeclared packages, fixed ports and 127.0.0.1 in containers, SQLite and disk writes on serverless, Prisma client not generated, Node version and lockfile mismatches

Deepest coverage: Next.js, React/Vite, Supabase, Firebase, Stripe, OpenAI/Anthropic/Gemini, Express, FastAPI/Flask/Django, Expo , with Midtrans and Xendit for Indonesian payments. The generic checks work on any stack.

Install

Claude Code (recommended: includes the guard hook)

/plugin marketplace add ribato22/vibecheck /plugin install vibecheck@vibecheck

Cursor, Codex, Gemini CLI, OpenCode, Windsurf and other agents (skills only)

npx skills add ribato22/vibecheck

Manual

Copy the folders in skills/ into ~/.claude/skills/ (or your agent's skills folder).

Commands

Command What it does

/vibecheck:check Full audit → score, verdict and .vibecheck/report.md with fix prompts

/vibecheck:check quick 2-minute pre-deploy check: secrets, database access and criticals

/vibecheck:check security Security areas only

/vibecheck:check launch Launch checklist (legal, SEO, monitoring) plus any critical

/vibecheck:fix Fix criticals, then highs, one at a time, with before → after score

/vibecheck:fix VC-DB-NO-RLS Fix one specific finding

/vibecheck:deploy Will it run in production? Env vars vs the host, localhost URLs, case-sensitive imports, ports, serverless storage, plus the host checklist → .vibecheck/deploy.md

/vibecheck:deploy --host vercel Same, with host-specific checks (vercel, netlify, cloudflare, railway, render, fly, heroku, docker, or local for apps that only run on your machine)

/vibecheck:unstuck Stop the fix loop: reproduce, find the root cause, revert the pile

You don't have to remember them. Ask "is my app ready to launch?" , "why does it work locally but not on Vercel?" or say "it's still broken, you keep going in circles" and the right skill kicks in. In other agents, call the skill by name ( check , deploy , fix , unstuck ).

The guard (always on)

With the Claude Code plugin, vibecheck watches every file write and every git commit / git push your agent makes:

vibecheck guard blocked this: it would put secrets where they do not belong.

• OpenAI API key sk-p…9xQz in src/lib/ai.ts:3
• NEXTPUBLICOPENAIAPIKEY has a browser-public prefix, so its value ships to every visitor (.env:2)

Fix: 1. Put the secret in .env (gitignored) and read it with process.env.NAME in server-only code. 2. Keep browser-public prefixes (NEXTPUBLIC, VITE, EXPOPUBLIC_...) for values that are safe to publish.

The agent reads that message and fixes it itself. The guard blocks only high-confidence key formats, allows secrets in .env , and checks every unpushed commit (not just the last one). Silence a false positive with a vibecheck-ignore comment on the line, or turn the guard off with VIBECHECK_GUARD=off .

Use it in CI

The scanner is a single Node file with no dependencies:

node skills/check/scripts/scan.mjs ./my-app # human-readable node skills/check/scripts/scan.mjs ./my-app --json # for tools and agents node skills/check/scripts/scan.mjs . --fail-on critical # exit 1 on high-confidence criticals

.github/workflows/vibecheck.yml

name : vibecheck on : [push, pull_request] jobs : scan : runs-on : ubuntu-latest steps : - uses : actions/checkout@v5 with : { fetch-depth: 0 } # full history, so .env files committed in the past are caught - uses : actions/checkout@v5 with : { repository: ribato22/vibecheck, path: .vibecheck-tool } - uses : actions/setup-node@v5 with : { node-version: 22 } - run : node .vibecheck-tool/skills/check/scripts/scan.mjs . --fail-on critical

Passed? Show it

Got a 🟢 GO? Add the badge to your README:

[ ![ vibecheck: GO 92/100 ] ( https://img.shields.io/badge/vibecheck-GO%2092%2F100-brightgreen )] ( https://github.com/ribato22/vibecheck )

Only GO apps get a badge. A public "NO-GO" badge would tell attackers where to look.

Safe by design

A security tool should never be the security problem.

The audit is read-only. It writes only to .vibecheck/ , which ignores itself in git. Fixes happen only in /vibecheck:fix , after you agree.

Zero dependencies, zero network calls , and it never runs your project's code. About 1,000 lines of plain JavaScript you can read in 15 minutes: rules.mjs , scan.mjs , guard.mjs .

Secrets are always masked ( sk-p…9xQz ) in output, reports and guard messages.

The guard fails open. If it crashes, your work continues.

Found a vulnerability in vibecheck itself? Please report it privately: see SECURITY.md .

FAQ

Does this replace a penetration test? No. It catches the mistakes AI tools make most often, and that's most real-world vibe-coding breaches. If you handle health, financial or children's data, get a professional review too.

Does it send my code anywhere? The scanner doesn't touch the network. The skills run inside the AI agent you already use, like any other prompt.

It flagged something that's fine. Add // vibecheck-ignore (or # vibecheck-ignore ) on that line. Put vibecheck-ignore-file in the first lines to skip a whole file, such as test fixtures. Then open an issue so the rule gets better.

My app was built in Lovable / Bolt / v0 / Replit. Export or sync it to GitHub, open it with any supported agent, and run the check. Paste the fix prompts back into your builder.

Contributing

New rules, stack guides and false-positive reports are very welcome. See CONTRIBUTING.md . Each rule ships with a test. Run them with npm test (Node 18+, no install needed).

Roadmap

Official GitHub Action

More stack guides: Convex, Clerk, Auth.js, Better Auth, Laravel, Rails

Live checks (security headers, open Supabase tables) behind an explicit opt-in

If vibecheck caught something before your users did, a ⭐ helps other builders find it.

MIT © vibecheck contributors · Not affiliated with Anthropic, Supabase, Firebase, Stripe or any other vendor named here.

About

🚦 Pre-launch security audit for AI-built apps. Catches leaked API keys, missing Supabase RLS, open Firebase rules and billing abuse. Launch Score + GO/NO-GO, guided fixes, secret guard. Claude Code plugin & Agent Sk

... (truncated at the site's per-page length limit; see the source link below for the full text.)