Vibe-coded apps ship with the service_role key in the browser, no row-level security, and the OpenAI key in the bundle. Anyone with devtools reads every customer's data. I move trust from the browser to Postgres + the server — then deploy it safely.
A representative AI feedback-triage SaaS, exactly as it comes out of Lovable/Bolt. Each row is a real issue and the exact fix. Left is what shipped; right is what I replace it with.
The client is built with the service_role key “so RLS doesn’t get in the way.” It bypasses all security — anyone who opens devtools has full read/write on every table.
Browser uses the anon key only. service_role is removed from the bundle, used solely in trusted server jobs, and the exposed key is rotated.
VITE_OPENAI_API_KEY + dangerouslyAllowBrowser. The key is extractable from the JS bundle — strangers spend on your account.
The LLM call moves to a server route /api/triage. The key is a server-only env var and never reaches the browser.
Tables are world-readable with the anon key; tenant isolation is a .filter() in React. One request returns every workspace’s tickets.
RLS enabled on every table with per-workspace policies via auth.uid() and a workspace_members table. Postgres refuses cross-tenant rows.
if (!user) navigate('/login') in a component. The data layer is wide open — skip the UI and the API answers anyway.
Auth enforced in middleware + server components; getUser() re-validates the JWT before anything protected renders.
isAdmin = localStorage.role === 'admin'. Any user types one line in the console and grants themselves admin.
Roles live in workspace_members.role and are enforced inside RLS (is_admin()), never trusted from the client.
User content rendered with dangerouslySetInnerHTML — a crafted ticket runs script in every teammate’s browser.
Rendered as plain text; React escapes it. User input is never treated as markup.
Raw user text dropped straight into the prompt, no length or type checks — trivially hijacks the model.
Input validated with zod; user text passed as delimited data with an “ignore instructions” guard; model output schema-checked.
The AI endpoint has no limit — a loop is a cost bomb. stripe_customer_id sits in a public-readable table.
Per-user rate limit / spend cap on the endpoint; billing moved to a table with no client policy (server-only access).
Every rescue ships with this checklist completed against your export — plus a second-account test that proves cross-tenant isolation actually holds.
Async, no calls required — I work from your export and a written walkthrough.
Share your Lovable/Bolt/v0 export or repo. I return a prioritized issue list — the same table you see above, against your code.
RLS migrations, server routes, secrets moved & rotated, validation, rate limits — wired to your Supabase project.
Second-account test proving cross-tenant isolation, dev/prod split, and a recorded walkthrough of every fix. Then it ships.
Open your deployed app, hit devtools → Sources, and search for “service_role”. If it’s there, you have a breach waiting to happen. Send it over — I’ll tell you exactly what’s exposed, free.