Opening a shop is easy: you rent the premises, set up the shop window, stock the shelves and open the doors. What you can’t see from the street is whether the storeroom door locks, and you don’t find out until someone walks through it. Something similar is what German consultancy Crissy Field found when reviewing apps built with Lovable, one of the trendy platforms for creating web apps by describing them in natural language.
The report, produced together with the newspaper Handelsblatt (which published its article on November 2), leaves little room for interpretation: many of those apps leave highly sensitive data within anyone’s reach, without needing to log in.
The numbers
They started from Common Crawl, the open web archive, and found 159,041 Lovable projects published between December 2025 and September 2026. They focused on the 34,545 using Supabase as their database:
| What | How many |
|---|---|
| Lovable projects analysed | 159,041 |
| Using Supabase | 34,545 |
| Loading database records on page load, without login | 24,664 |
| With tables containing email addresses | 7,403 (10,560 tables) |
They’re careful with the numbers themselves: a page loading data without login isn’t a flaw in itself (it may be public data), some emails will be contact or test addresses, and that 30% isn’t a confirmed leak rate. But then come the examples, and there caution doesn’t help much:
- A German forum with over 80,000 profiles and more than 200,000 chat messages.
- Event organisers and restaurants with tens of thousands of sales, including phone numbers, addresses, payment data and ticket codes.
- A service provider with over 1,000 identity document records: passport numbers, expiry dates and nationalities.
- A property management company with AI-generated summaries of internal emails about tenants and owners.
- Contracts, invoices, personal photo albums, internal WhatsApp conversations and 280,000 emails sent to a provider’s customers.
To check, they made read-only requests with the same key each site used and only downloaded parts of the data. They don’t publish the access details, and they couldn’t determine whether anyone else had exploited them.
Where the problem lies
The report doesn’t go into the exact technical cause; it only says database integrations follow “recurring patterns”, which makes it very easy to look for the same mistake across thousands of apps at once. But to understand why this happens, it helps to know how Supabase works: the browser talks directly to the database using a key that is public (it’s in the page’s code, visible to anyone), and the only thing stopping anyone from reading any table is the access policies defined on each one. If a table doesn’t have them, or they’re wrong, the storeroom door is open.
It’s a model that works very well when you know it exists. The problem is that the whole point of Lovable is that you don’t need to know: you describe what you want, the tool builds it, and the app works. It works perfectly, in fact, and that’s the trap. An app that lets anyone read every user’s data throws no errors, doesn’t look odd, and passes any test done by someone who only checks whether it “works”.
The unsupervised intern
This is exactly what I was talking about in the watercolour article: AI is a great intern, but it needs supervising, and to supervise it you need to know the guts of what it does. Someone who doesn’t know what access control is can’t notice one is missing. And it’s not just a problem for people who don’t code: a few days ago I wrote about how Claude Code stored secrets in a file that ended up on GitHub, and that happens to people who know perfectly well what a secret is.
The report’s recommendations are the usual ones, and no less true for it: review permissions, interfaces and separation between user accounts before putting real data in, repeat the review after every significant change and, if there’s no expertise in-house, get it from outside. They sum it up in one line themselves: automated checks are no substitute for a professional security assessment.
Building an app has never been easier, and that’s great. But how easy it is to open the shop doesn’t change what happens if the storeroom door is left open, and the people whose passports were in those tables didn’t choose which tool was used to build it.




