Your vibe-coded app is leaking: what the Lovable security breach means for non-technical founders

On April 20, 2026, security researchers disclosed that Lovable — the AI app builder — had left every project created before November 2025 reachable by any free-tier account. Source code, database credentials, AI chat history, and customer data. A researcher reported the flaw to Lovable via HackerOne on March 3. Lovable patched new projects but left older ones live for 48 days. When researchers went public Monday, Lovable's first public response called the exposure "intentional behavior" before walking it back.

One audited Lovable-built app contained 16 vulnerabilities. Six critical. Roughly 18,000 users' records were exposed.

If your business runs on an app built with Lovable, v0, Bolt, Cursor, or any other AI app builder, stop reading and rotate your keys. Then come back.

What happened

The flaw is a Broken Object Level Authorization (BOLA) on Lovable's platform API. Any authenticated user could enumerate other users' projects and pull the contents. The exposed surface covered source code, Supabase connection strings, AI chat history (including any secrets pasted into prompts), and customer records from whatever database the generated app connected to.

Researchers reported the vulnerability on March 3, 2026. Lovable rolled out a patch for projects created on or after November 2025 on April 2. Projects created before that date stayed exposed until the researchers disclosed publicly on April 20. That is a 48-day window where the flaw was known to Lovable, patched for new customers, and left live for existing ones. Accounts tied to employees at Nvidia, Microsoft, Uber, and Spotify were among those affected.

The compounding issue: Lovable ships apps that connect to Supabase without row-level security policies enabled by default. Supabase with RLS off will serve any row to any authenticated request. So even apps on patched infrastructure leak at the application layer if the person who prompted them didn't know to turn RLS on. Most non-technical founders don't.

Why every AI app builder has this problem

The studio has audited apps generated by Lovable, v0, Bolt, and Cursor on behalf of founders. They share a pattern.

• Authorization is wired at the UI layer (hidden buttons, React guards). Not at the API. The server trusts whatever the client asks for.
• Supabase or Firebase connections ship with admin-level keys in the client bundle.
• Row-level security policies either don't exist or permit everything.
• Environment variables with real production secrets sit in the generated repo.
• Input validation lives on the happy path. The LLM does not think adversarially.

None of this is the AI's fault. These platforms optimize for time-to-first-demo, not time-to-safely-deployed. Founders who have never shipped production software don't know what's missing. They find out when something leaks.

What should I do this week?

Three actions. Today.

1. Rotate every credential the app has touched. API keys for Supabase, Firebase, Stripe, OpenAI, SendGrid, Twilio — assume exposed. If you built on Lovable before November 2025, assume your chat history was pulled, and any secret you pasted into an AI prompt is compromised.

2. Turn on row-level security. In Supabase: RLS on for every table, default-deny, then write explicit policies per table. In Firebase: tighten your security rules. In any custom backend: verify authorization runs server-side, not only in the UI.

3. Have a developer read the generated code. Not the product. The code. Every route handler. Every database query. Every auth check. Grep for embedded secrets. Verify admin routes are gated on the server.

If you don't have a developer who can do all three, pause the app and get an audit this week.

What an audit of an AI-built app looks like

A Kingbird audit of an AI-generated app is targeted at the failure patterns AI builders produce. It is not a generic security review.

The process runs two to five business days for a typical Lovable-sized app:

• Enumerate every API endpoint. Test each one without auth, with a different user's token, and with malformed input.
• Read every database policy. Compare it against what the UI implies the user can do.
• Grep the generated repo for embedded secrets.
• Verify admin routes are gated server-side, not only in the UI.
• Review the LLM's chat history for anything that should never have been committed to the repo.

The deliverable is a written list of what is exposed, what needs to change in the next 30 days, and a runbook for turning on the security controls the AI skipped. If the app is salvageable, the studio patches and hardens. If the auth layer, data access layer, or payments code can't be repaired, we rebuild those pieces.

Who we are

Kingbird Solutions is a custom software development studio based in Phoenix, Arizona. The studio takes over stalled software products, runs managed development retainers, and ships fixed-scope builds for non-technical founders. A senior US operator, Chris King, is on delivery from kickoff through handoff on every engagement.

The studio built and ships BuddySOS, a consumer safety app with real customer data and real security constraints. The audit process above is the same one we run on our own code before a release.

Next steps

Five audit slots open this week. Pick the path that fits.

• You want a fast read on whether your AI-built app is leaking. Run the Stalled Project Diagnostic. Seven questions, five minutes, a real answer. No signup wall.
• You want the studio to audit the code this week. Book a 30-minute fit call. We'll tell you what an audit costs, what we'll likely find, and what comes after.
• You already know you need an audit and want a fixed-price proposal. Book a scoping call with AJ. Fastest path to a contract.
• You want a paid, low-risk review before committing. The Stabilization Review is a ten-business-day engagement at $4,500 flat. Written 30/60/90 plan plus a live debrief. You keep the plan whether or not we keep working together.

Whatever path you pick, do not leave a Lovable, v0, or Bolt app in production without an engineer reading the code. This week.