SecurityAI-built appsGuide · Updated Aug 2026

Vibe coding security: what actually goes wrong

Apps built by prompting ship a recognisable set of holes — not because the model is careless, but because nobody asked it the boring questions. This is the list of what really breaks, and the checklist that closes each one.

Why AI-built apps share the same holes

A coding agent optimises for the demo working. Security is invisible in a demo: the app with RLS off renders exactly like the app with RLS on — until someone else’s data shows up. So the same failure modes repeat across Bolt, Lovable, v0, Cursor and Claude-built apps, and they are worth naming precisely, because every one of them is checkable.

The hardening checklist, by platform

Supabase

Vercel & the frontend

GitHub / GitLab

Cloudflare

Make it continuous, not a launch-day ritual

Every item above drifts. The table added next sprint ships without RLS; the debug flag comes back with a hotfix; the new function skips the auth check. A checklist you run once is a snapshot — the failure mode is everything that changes after it. That is the entire reason lumioguard exists: it runs 323 checks like these against your repo and your live services on every scan, and the fixes arrive as pull requests you review, not tickets you triage.

Run the checks on your app

Connect your repo and live services read-only. First scan is free — see every one of these findings on your own app in minutes.