| name | vibe-audit-technique |
| description | White-box auditing methodology for AI-generated ('vibe-coded') applications from Lovable, Bolt.new, v0, Cursor, Replit Agent, and Claude Code. Focuses on modern stack misconfigurations (Supabase, Firebase, Next.js, Vercel, Expo). |
vibe-audit-technique
Goal: Identify critical, high-frequency security omissions in modern web/mobile stacks (Next.js, Firebase, Supabase, Vite, Expo) that are typically generated by AI coding assistants (Lovable, Bolt.new, v0, Cursor, Replit Agent, Claude Code).
When this technique applies
- You have source-code access to a modern web application or mobile app backend.
- The stack relies heavily on BaaS (Backend-as-a-Service) like Supabase, Firebase, or Clerk.
- You are reviewing Next.js Server Actions, tRPC routers, React Server Components, or Expo/React Native apps.
- The repo shows AI-generator fingerprints (Lovable/Bolt/v0/Cursor commit messages, one-shot scaffolded migrations,
.cursor/, mcp.json, .bolt/).
Pair with sibling techniques
- For LLM-assisted taint tracing from source to sink across large repos, drive the pass with
offensive-techniques/source-review-technique/SKILL.md. This skill supplies the vibe-code sink catalogue; that skill supplies the AI-assisted data-flow discipline.
- Do not re-derive general SAST here.
The Audit Workflow
1. The BaaS Access Control Trap
AI code generators frequently scaffold databases without strict Row-Level Security (RLS). This is the canonical vibe-code failure — see CVE-2025-48757 (Lovable + Supabase, mass RLS-off exposure across hundreds of live apps).
- Check: Are tables missing RLS entirely (default state for
public schema tables in Supabase)?
- Check: Are existing RLS policies set to
USING (true) or USING (auth.uid() IS NOT NULL) for convenience?
- Check: Cross-tenant IDOR — do policies scope to the row owner / organization, or only to "any authenticated user"?
- Action: Load
references/database-security.md. If ORM (Prisma/Drizzle) or raw SQL is layered on top, also load references/data-access.md.
2. The Client-Side Secret Leak
AI often pushes server-side secrets to the client to quickly bypass CORS or backend proxy errors.
- Check: Grep for
NEXT_PUBLIC_, VITE_, or EXPO_PUBLIC_.
- Check: Ensure OpenAI, Anthropic, or Stripe secret keys are not bundled in these variables.
- Action: Load
references/secrets-and-env.md and references/ai-integration.md.
3. Server Action / API Auth Bypass
In Next.js, Server Actions are hidden API endpoints. They must be authenticated individually.
- Check: Does the endpoint or Server Action assume the user is trusted simply because middleware exists? Middleware is often bypassed or overly permissive.
- Action: Load
references/authentication.md.
4. Billing and Rate Limits
AI assistants build features, not defenses.
- Check: Are AI query routes (e.g., text generation, image generation) protected by rate limits?
- Check: Are payment webhooks properly verifying signatures?
- Action: Load
references/rate-limiting.md and references/payments.md.
5. Mobile Bundle & Deployment Surface
Expo/React Native apps generated by AI ship with embedded secrets and insecure storage. Vercel/Netlify preview and prod configs are often mixed.
- Check: Any
EXPO_PUBLIC_* referencing a secret key; use of AsyncStorage for tokens; deep links used for auth.
- Check: Preview deployments sharing production credentials; source maps or
.git/ exposed in prod; permissive CORS.
- Action: Load
references/mobile.md for the bundle surface, references/deployment.md for the hosting surface.
Quality Gates
- Context matters: A leaked
NEXT_PUBLIC_STRIPE_PUBLISHABLE_KEY is not a vulnerability (it is meant to be public). A leaked STRIPE_SECRET_KEY is critical. Verify before reporting.
- Do not report generic 'missing headers': Focus strictly on the critical gaps caused by "vibe-coding" (Direct DB exposure, unauthorized server actions, secret leakage).
