7,403 Lovable Apps Hand Out Their Databases. No Login Needed.

Crissy Field scanned 159,041 Lovable projects and found 7,403 serving user emails without a login. Lovable's head of security says better tooling is the fix, while the terms put data handling on the customer.

7,403 Lovable Apps Hand Out Their Databases. No Login Needed.

7,403 apps built on Lovable will hand you their users' email addresses without a login. Some will hand you a lot more than that.

The number comes from Crissy Field, a security firm in Hamburg, working with the business paper Handelsblatt. Using the open web archive Common Crawl, they identified 159,041 Lovable projects built between December 2025 and September 2026. Of those, 34,545 used Supabase, the database service behind most Lovable projects. 24,664 sent database records the moment the page loaded, before anyone signed in. Serving records to a logged-out visitor is sometimes by design, so the team made additional read-only requests with the same anonymous key. Email addresses turned up in 10,560 tables across 7,403 sites, 30 percent of the ones they probed. On the Hacker News thread, the lead researcher added a detail the report leaves out: a screening pass with an 85 percent confidence threshold flagged close to 4,000 sites as leaking privacy-sensitive content.

The case list covers everything people ship in practice. A German lifestyle forum exposed more than 80,000 profiles with full names and birthdates, plus over 200,000 private chat messages. Event organizers and restaurants leaked tens of thousands of ticket and food-order records, complete with payment details and ticket codes. A property management company exposed AI-generated summaries of internal emails about tenants and business partners. One service provider left over 1,000 identity document records readable: passport numbers, expiration dates, nationalities. Elsewhere the researchers found contracts, invoices, personal photo albums, internal WhatsApp threads, and in one case 280,000 emails a provider had sent its customers. Crissy Field could not establish that any of it was misused. They also only pulled parts of each dataset, so the true counts are likely higher.

The mechanism is dull, which is why it scales. Supabase ships every frontend an anonymous API key, and Row-Level Security policies decide what that key can read. Lovable's generator writes those policies, and it writes them open. The people Lovable markets to, the ones who need no technical knowledge, have no way to reason about RLS. The report adds a sharper point: database integrations on the platform follow recurring patterns, so an approach that works against one app likely works against thousands. Lovable's terms put responsibility for applications and data handling on the customer. The marketing says you need to know nothing. The fine print says otherwise.

Lovable's head of security showed up in the thread on November 14. Igor confirmed the root cause in one line: insufficient Row-Level Security. His answer is tooling. Every app is scanned before publish, free for all users, a deeper SAST scan is one click away, flagged issues ship with a Fix button, and auto-fix resolves findings in the background. Publishing can be blocked entirely when critical findings exist. Commenters pushed back on that arrangement: with 7,403 leaking sites, warnings nobody reads look like a systematic failure, and a platform selling "no technical knowledge needed" should block leaky apps outright instead of warning them. Igor's reply is worth reading, and his security post lays out the same stack. Users fear fixes that break production, so his team is training auto-fix to trace changes back to user intent, with the goal of switching it on by default. Soon, he says.

One side plot matters to the agent crowd. When commenters challenged the ethics of pulling real records, the researcher said a lawyer was involved from the newspaper side and every LLM pass ran locally, an "elephant-1" model on a single RTX Pro 4000 SFF. The leaked records never touched a cloud API. If you audit systems you do not own, that detail is the standard to copy.

If you ship an AI-built app with real users, run the researchers' first check on your own product: open it in a private window and watch what the API returns before any login. Then get the access rules reviewed by someone who reads RLS policies for a living. Crissy Field closes the report plainly. Automated checks are no substitute for a professional security assessment before an app touches real customer data.

This is last week's story one layer up. Claude Code wrote API keys to local files that landed in public GitHub repos, so the desktop agent leaked secrets by default. Here the platform shipped open databases by default and the terms called the customer responsible. Different layer, same shape: AI tooling optimizes for shipping, and the insecure default stays in place until somebody enforces the secure one.

Igor says auto-fix on by default is coming. When it ships, compare the dates in Lovable's changelog against the next independent scan. That comparison, not the blog post, will tell you whether the defaults moved.