AI-Built Apps Leak Passports, Contracts and 200,000 Chats in Massive Supabase Exposure

A large-scale security audit of applications built with the AI app-generation platform Lovable has uncovered thousands of databases openly serving sensitive user records to anyone who asks — no login required. The findings, published by German application-security firm Crissy Field in collaboration with business daily Handelsblatt, paint a stark picture of the security debt accumulating behind the low-code and “vibe coding” boom.

  • Scale of the exposure: Researchers scanned 159,041 Lovable projects and found 24,664 Supabase-backed sites returning database records without any authentication.
  • What leaked: More than 1,000 passport records with nationalities and expiry dates, tens of thousands of contracts and invoices, 7,403 sites with readable email addresses, and a forum exposing over 200,000 chat messages and 80,000 user profiles.
  • Root cause: Missing Row-Level Security (RLS) rules in Supabase, not a breach of Lovable’s own infrastructure. The public anon key shipped in every app’s frontend was enough to read entire tables.

How the audit was conducted

Crissy Field pulled its sample from the public Common Crawl web archive, focusing on Lovable projects created between December 2025 and September 2026. Of the 159,041 projects identified, roughly 34,545 relied on Supabase as their backend — the hosted Postgres service that Lovable uses almost by default. From that subset, 24,664 sites returned database records immediately upon page load, before any credential was supplied.

The researchers then repeated read-only requests using the same public API key the site had already handed to the browser. No exploitation, no brute force, no breach of Lovable’s servers. If the data came back to the researchers, it would come back to anyone who looked.

Among the specific findings: email addresses were readable across 10,560 tables on 7,403 sites. A German lifestyle forum leaked more than 80,000 profiles containing full names, email addresses, birth dates and gender, alongside over 200,000 chat messages. Event organizers and restaurants exposed tens of thousands of ticket and food sales records with phone numbers, postal addresses, payment information and ticket codes. A property management firm had AI-generated summaries of internal emails — including tenant and landlord data — sitting in the open. One CRM-style project exposed 280,000 customer emails, waiting-list entries and support tickets.

The technical mechanism behind the leaks

The exposure stems from how Supabase handles access control. Every Lovable app communicates with its database through a public key called the anon key, which is designed to be visible in frontend code. The only barrier between that key and the underlying data is Row-Level Security — a set of database rules dictating which rows a given user may read. Supabase’s own documentation is explicit that enabling and writing those rules is the developer’s responsibility.

When RLS is switched off, Supabase does exactly what it is told: it returns every row to anyone holding the public key. That single misconfiguration explains every number in the report. A sharper variant of the same mistake involves shipping the service_role key — an administrative credential that bypasses RLS entirely — in frontend code, which leaves a database wide open regardless of policy settings.

Lovable maintains that the issue is a known and common Supabase configuration problem rather than a flaw in its platform, and its terms place application security with the customer.

The company has pointed to the industry-wide nature of the problem. Independent scans have catalogued more than 16,000 exposed Supabase databases behind vibe-coded apps, and an earlier critical-severity vulnerability, CVE-2025-48757, described precisely this class of insufficient RLS in Lovable-generated sites. To its credit, Lovable responded to that CVE by adding an automated Security Scanner and an AI Security Reviewer — though researchers note that checking whether a policy exists is not the same as checking whether it is correct.

Why AI builders concentrate the risk

None of this is unique to Lovable. The same gap appears in apps built with Bubble, Replit, Base44, v0 and hand-written code. But AI builders amplify the problem for structural reasons. Speed hides steps: a developer describes a feature, tables appear, the app works, and the invisible security layer is easy to assume is handled. Generating a working app and generating a secured one are different jobs, and the quickest path to “it works” leaves RLS off. Traditional teams have someone who asks whether a table is protected; a solo founder shipping with AI often does not — until a researcher, or an attacker, asks on their behalf.

The researchers caution that the widely cited 30% figure is not a confirmed leak rate. It is calculated only against the 24,664 sites that returned data on page load, not the full 159,041 projects, and the Common Crawl sample is not representative of all Lovable apps. Some exposed tables may hold public-facing data such as product catalogues. Still, the pattern is consistent, and the fix is well documented.

The broader lesson is that the promise of vibe coding — anyone can ship an app — is real, but an app is not only what is visible on screen. It is also what it stores. Until platforms configure security correctly by default, every improvised application remains a potential leak waiting to happen, and the responsibility cannot rest solely with builders who never knew what Row-Level Security was in the first place.

Sources & References