Lovable AI Apps Expose Passports and 200,000 Private Messages

Open database doors across 7,400 sites

A systematic crawl of Lovable web applications revealed that thousands of databases allowed unauthenticated read requests.

In a collaborative investigation with German business outlet Handelsblatt, security consultancy Crissy Field audited 159,041 projects built on Lovable, the Swedish platform that converts natural language prompts into live web applications. The researchers pulled application links logged in Common Crawl between December 2025 and September 2026, focusing on deployments backed by Supabase.

Among the 34,545 apps using Supabase, 24,664 transmitted database records immediately upon page load without requiring authentication. When researchers issued basic read-only queries using the public API keys embedded in those front ends, they were able to extract tables containing email addresses across 7,403 sites, roughly 30% of that sample.

The problem does not stem from an infrastructure breach at Supabase itself, but rather a catastrophic failure of basic row-level security policies during automated code generation. Similar systemic oversights have emerged elsewhere as autonomous tools ship raw code into production, reminiscent of how AI coding agents leak secret keys when operating without guardrails.

From tenant files to passport numbers

The data uncovered during the audit went far beyond harmless test records. In one severe instance, a service provider left an open table containing more than 1,000 identity records, exposing full names, nationalities, expiration dates, and passport numbers.

A German lifestyle forum built on the service exposed 80,000 user profiles alongside more than 200,000 unencrypted chat logs. Event booking portals and dining apps leaked ticket codes, street addresses, and payment receipts. A property management tool openly served AI-generated email summaries that detailed private disputes involving tenants and building owners.

Because Lovable generates software using predictable application templates, the architectural flaws repeat across multiple sites. Once researchers identified how one application exposed its backend, querying identical misconfigurations on other Lovable instances required very little additional effort.

The liability trap of prompt-to-app development

Lovable was launched in late 2024 by founders Anton Osika and Fabian Hedin, evolving from the popular GPT Engineer project. Its central marketing pitch has consistently promised that anyone can construct functional software without understanding code or infrastructure. But behind that accessibility lies a sharp legal contradiction.

Lovable's terms of service place complete responsibility for data protection and security management squarely on the account holder. A beginner who prompts an application into existence is legally expected to audit database permissions, authentication headers, and row-level access control, tasks they fundamentally lack the training to understand.

While researchers noted there is no evidence this specific cache of data was stolen by malicious actors, the exposure illustrates a growing hazard. When platforms abstract away technical complexity, they frequently disguise the maintenance and hardening steps necessary to keep user data private.

What it means for you

If you run an application built on Lovable that collects form responses, user registrations, or customer messages, check your Supabase dashboard immediately. Ensure that Row Level Security (RLS) is enabled on all tables containing non-public data, and verify that the anon API key cannot read entire collections.

End users who have entered personal details, booking information, or identification documents into small, newly launched web services should monitor their inboxes for targeted phishing attempts. Lovable has not yet released automated migration tooling to retroactively lock down previously deployed client databases.