You describe an app in plain English, the AI builds it, you ship it. That’s the pitch of Lovable, the Swedish darling of vibe coding. What the pitch leaves out is who’s supposed to lock the database. A German security firm just answered that question by looking at a lot of Lovable apps, and the answer is: nobody.
Crissy Field, a Berlin-based application security outfit, pulled 159,041 Lovable projects out of Common Crawl archives covering December 2025 to September 2026. Of those, 34,545 used Supabase as their backend, and 24,664 of them happily sent database records to any visitor the moment the page loaded. No login, no interaction, nothing.
What the Lovable apps were leaking
Sending some data on page load isn’t a crime by itself; maybe it’s a public list of products. The problem is what happened when the researchers asked the same API, with the same public key, for a little more. In 7,403 of those sites (30%), they found tables full of email addresses, which is a reliable sign of personal records sitting in the open.
Then come the examples, and they’re ugly. A German lifestyle forum exposed over 80,000 profiles with full names, birth dates and gender, plus more than 200,000 private chat messages. Restaurants and event organisers leaked tens of thousands of ticket and food orders with phone numbers, postal addresses and payment details. A property manager exposed AI-generated summaries of internal emails about tenants and owners. And one service provider left more than 1,000 identity records readable, passport numbers and expiry dates included.
Add contracts, invoices, personal photo albums, WhatsApp exports and, in one case, 280,000 emails sent to a company’s customers. The full research report is careful to say the 30% figure isn’t a confirmed leak rate and that they only sampled the data, which honestly makes it scarier, not less.
Why vibe coding keeps failing the same way
None of this is an exotic hack. It’s the classic Supabase row-level security story: the anonymous key is meant to be public, and your tables are only as private as the policies you wrote. If you never wrote any (and you don’t know what a policy is, because the marketing told you no technical knowledge is needed), everything is readable.
Lovable’s terms put responsibility for data handling squarely on the customer. Fair enough legally, but there’s a gap the size of a GDPR fine between “anyone can build an app” and “you’re on your own for security”. And because every Lovable app wires up its backend the same way, one trick that works on one app works on thousands. The researchers point this out explicitly, and it’s the same pattern we saw with Claude Code sessions leaking API keys to public GitHub: AI tools don’t create new mistakes, they industrialise the old ones.
What I’d do if I had a Lovable app in production
First, stop assuming the platform did it for you. Open your Supabase dashboard, check that RLS is enabled on every table that holds anything about a human being, and actually read the policies. Second, test your own app logged out, in a private window, and see what the network tab shows. If you see rows you wouldn’t want on a billboard, you have a problem today, not a hypothetical one.
Third, if the app handles payments, IDs or health data, pay a human to review it before more customers trust you with it. That’s not a sales pitch for consultants, it’s the cheapest incident response you’ll ever buy. And keep a copy of your code and data somewhere you control: Mozilla Solo users learned this month what happens when an AI site builder decides it’s done hosting you.
Handelsblatt, which worked with the researchers, has asked Lovable for comment. I’d love to hear the answer, but I suspect it’ll sound a lot like “security is a shared responsibility”, which is what every platform says right before your users’ passport numbers end up in a Telegram channel.
Vibe coding gets you an app in an afternoon. Turns out it also gets everyone else one.
Discover more from TechPassion
Subscribe to get the latest posts sent to your email.





