After scanning nearly 160,000 apps built with the AI coding platform Lovable AI, researchers found passport numbers, private chat messages, and payment details in databases that didn’t even need a simple login to open.

Lovable AI CEO Anton Osika confidently told CNBC this week that AI was collapsing barriers to starting a business.

However, it seems that his company – the flagship of vibe coding since it turns a sentence typed in plain language into a working web application – has also collapsed protections around the customers’ personal and confidential information.

Researchers from a German security firm, Crissy Field, combed through 159,041 projects created on Lovable AI and found passport numbers, payment information, and more than 200,000 chat messages reachable straight from the browser.

Highly sensitive details in the open

Their idea was to pull these projects from the Common Crawl web archive and see exactly what their databases would hand over to a stranger with no credentials whatsoever.

The answer, as Crissy Field researchers detail in their write-up of the research, is devastating.

Of the 34,545 projects running on Supabase (the service Lovable AI almost always uses to store the data of the apps it generates), 24,664 returned database records with zero authentication required.

According to the researchers, email addresses alone showed up readable on 7,403 sites, spread across more than 10,500 tables.

Moreover, one service provider’s database shelled out over 1,000 passport records complete with expiration dates and nationalities. These highly sensitive identity details could be used for identity theft.

Event organizers and restaurants let anyone read tens of thousands of orders and tickets sold, with emails, phone numbers, postal addresses, payment information, and ticket codes. Again, this is precisely the raw material for a scam at the door.

A lifestyle forum built on Lovable AI leaked more than 200,000 chat messages and 80,000 user profiles – to anyone who asked.

A property management company exposed AI-generated summaries of internal emails containing sensitive information about tenants, property owners, and business partners.

Other accessible content included tens of thousands of contracts, invoices, personal photo albums, and internal WhatsApp communications.

Crucially, nobody forced anything. Researchers only used the access those sites offer to every visitor, with read-only requests and the same key the page hands to the browser.

Industry-wide issue with Supabase?

“Our analysis shows that many Lovable applications leave even highly sensitive data unprotected, with potentially serious consequences for the people affected,” Crissy Field said.

“An easy-to-use development platform does not remove the need for an expert review of access permissions before an application handles confidential data.”

In other words, researchers think that because Lovable AI enables non-technical users to spin up full-stack apps quickly with plain-language prompts, crucial security reviews often get skipped.

The vulnerabilities seem to stem from unconfigured or missing Row Level Security rules in backend databases rather than a hack of the Lovable AI platform itself.

The scale of the problem is not insignificant. Lovable AI, a Swedish AI startup founded in Stockholm in 2023, reports more than 50 million projects already delivered and 1.2 million new ones per week.

However, the company responded to the researchers’ findings – also published in the German business daily Handelsblatt – by claiming that they’re dealing with a known, industry-wide configuration problem around the use of Supabase.

The security company UpGuard recently examined around 300,000 domains with indications of Supabase. On 16,326 databases, tables were readable from outside. According to UpGuard, more than half of these cases contained indications of personal data.