Follow us on Google News
Thousands of applications built using Lovable AI exposed sensitive user information through publicly accessible databases, including passport details, payment information, private chat messages, contracts and internal communications, according to research by German security firm Crissy Field.
The researchers analysed 159,041 projects associated with Lovable AI and found widespread cases in which backend databases could return records without requiring authentication.
Lovable AI is a platform that allows users to create web applications using natural-language prompts. Many of those applications rely on Supabase for backend database services.
The researchers said the problem was not caused by a breach of Lovable AI itself but by missing or improperly configured database access controls in applications built on the platform.
Thousands of Databases Returned Data Without Login
Crissy Field examined projects collected through the Common Crawl web archive and identified 34,545 Lovable applications using Supabase.
Of those, 24,664 returned database records without requiring authentication, according to the research.
The researchers then examined the exposed records using read-only requests and found email addresses accessible across 10,560 tables on 7,403 sites.
The testing did not involve bypassing authentication controls or exploiting a vulnerability in Lovable AI. Researchers said they used the same access mechanisms already available to ordinary visitors through the applications.
The scale of the exposure highlighted the risks created when database permissions are left open by default or are not reviewed before an application is placed online.
Passport and Identity Details Found Exposed
Some of the exposed databases contained highly sensitive personal information.
One service provider reportedly exposed more than 1,000 passport-related records containing details such as passport numbers, expiry dates and nationalities.
Such information can be valuable for identity theft, impersonation and fraud.
Other applications exposed data connected to event bookings, restaurant orders and ticket sales.
These records included combinations of email addresses, phone numbers, postal addresses, payment information and ticket codes.
The researchers warned that such information could be used for highly targeted scams because attackers would have access to both personal details and transaction-related context.
200,000 Private Chat Messages Left Accessible
One of the largest exposures involved a lifestyle forum built using Lovable AI.
Researchers found more than 200,000 chat messages and approximately 80,000 user profiles accessible through its backend database.
Other exposed content across different applications included contracts, invoices, photo collections and internal communications.
In one case, a property management company exposed AI-generated summaries of internal emails containing information about tenants, property owners and business partners.
Internal WhatsApp communications were also reportedly accessible in some applications.
The findings show that applications built rapidly with AI tools can end up handling significant amounts of confidential data even when their creators may have limited experience with backend security.
Researchers Point to Missing Supabase Access Rules
The researchers attributed the exposure primarily to missing or incorrectly configured Row Level Security rules in Supabase.
Row Level Security can be used to restrict which database records individual users are allowed to access.
If these controls are not configured correctly, applications may unintentionally allow unauthenticated visitors to query database tables directly.
Crissy Field said an easy-to-use AI development platform does not remove the need for expert review of permissions before an application begins processing sensitive information.
The issue therefore appears to be less about an attacker breaking into systems and more about applications being deployed with data-access rules that were too permissive.
Lovable Says Issue Extends Beyond Its Platform
Lovable AI responded by describing the findings as part of a broader configuration problem affecting applications built with Supabase rather than a breach unique to Lovable.
That position is supported by separate research into Supabase deployments.
Security company UpGuard recently analysed around 300,000 domains showing signs of Supabase usage and found 16,326 databases where tables were readable externally.
According to that research, more than half of those exposed databases showed indications that they contained personal information.
The findings suggest that insecure Supabase configuration can affect applications created through traditional development methods as well as those generated through AI-assisted coding tools.
Vibe Coding Raises New Security Questions
Lovable AI has become one of the most prominent examples of so-called vibe coding, where users describe an application in plain language and an AI system generates much of the underlying code.
The approach significantly lowers the technical barrier to building software.
However, it can also allow people with limited experience in application security, authentication and database permissions to deploy full-stack services that collect sensitive information.
The Crissy Field findings show how quickly that risk can become serious when an application stores passports, payment information, communications or other confidential material.
AI-generated applications may appear functional from a user-interface perspective while still containing serious weaknesses in their backend configuration.
No Evidence of a Central Lovable AI Breach
Importantly, the research does not show that attackers compromised Lovable AI’s own infrastructure or stole data from a central Lovable database.
Instead, the exposed information came from individual applications built using the platform and connected to backend databases with insufficient access restrictions.
That distinction matters because describing the incident as a Lovable AI data breach would overstate what the research established.
The more accurate issue is widespread data exposure caused by insecure application and database configuration.
Security Reviews Remain Essential for AI-Built Apps
The research underscores the need for security testing before AI-generated applications are used to process personal or business information.
Developers should review database access policies, ensure Row Level Security is enabled where appropriate and test whether unauthenticated users can retrieve records that should remain private.
Applications handling identity documents, payment data, private communications or contracts require particularly strict controls.
The findings also raise a broader challenge for AI coding platforms: making software development easier does not automatically make the resulting applications secure.
As AI tools allow more people to build and deploy software quickly, database permissions, authentication rules and privacy controls may become some of the most important safeguards separating a working application from a serious data exposure incident.
About the author — Suvedita Nath is a science student with a growing interest in cybercrime and digital safety. She writes on online activity, cyber threats, and technology-driven risks. Her work focuses on clarity, accuracy, and public awareness.
