A German security firm Crissy Field and Handelsblatt joint investigation quantified an alarming pattern: of 159,041 apps generated by Lovable, 24,664 returned database records the instant their pages loaded—no login, no captcha, no interaction. This wasn't a hacker's work; the websites handed the data over themselves.
Lovable is a Stockholm-based company that turns a natural-language description into a working web app, marketed as requiring "no technical knowledge." Its backing database service is Supabase, where a single configuration setting determines which rows of which tables are visible to whom. In every case found, that setting was never switched on.
The Leak Path: Not a Single Complex Step
The entire chain involved no attack activity. Researchers filtered Lovable projects from the open web archive Common Crawl, visited each one, and recorded what was returned on first page load; for sites returning database records, they used the publicly visible API key already present in the browser to issue read-only queries.
| Step | Number | Note | |---|---|---| | Lovable projects identified via Common Crawl | 159,041 | Dec 2025 – Sep 2026 | | Of which use Supabase | 34,545 | The dominant database integration in the Lovable ecosystem | | Returned database records on first page load | 24,664 | No login or interaction required | | Records containing emails after additional read-only queries | 7,403 sites / 10,560 tables | ~30% of the former |
The report itself cautions readers against interpreting the 30% figure as a leak rate: emails may be public contact info or test data; personal records without emails would be missed, and these biases point in opposite directions. The sample comes from Common Crawl and doesn't represent all Lovable apps. Researchers only downloaded portions of datasets—the real numbers are likely larger.
What Leaked: Passport Numbers at the Top of the Severity Scale
- A German lifestyle forum: 80,000+ readable profiles with full names, emails, birth dates, genders, plus 200,000+ chat messages
- Several event organizers and restaurants: tens of thousands of ticketing and dining orders—emails, phones, addresses, payment details, ticket codes
- One service provider: 1,000+ ID document records including passport numbers, expiry dates, and nationalities
- Contracts, invoices, private photo albums, internal WhatsApp communications; one provider had 280,000 customer emails sitting in queryable tables
- A property management company leaked AI summaries of internal emails containing sensitive details on tenants, owners, and partners
Why an Attack-Free Failure Is Systemic
Row Level Security (RLS) is disabled by default in Supabase; configuring it is the developer's job. Hand-written projects vary widely—one misconfiguration leaks one app. Lovable's generated apps are highly uniform: apps from similar prompts share the same data-layer structure, so one missing rule can be copied into thousands of apps. The report's conclusion: one probe sweeps across a whole batch at almost no extra cost.
Lovable claims over 50 million delivered projects and 1.2 million new ones weekly. An independent UpGuard scan separately found 16,326 readable Supabase databases—the problem extends far beyond one platform. This is not one company's bug; it's the default outcome of a class of development patterns.
The responsibility line matters too: Lovable's marketing says no technical knowledge is needed, but its terms place all liability for apps and data processing on the customer. A company with no engineers signing those terms likely doesn't know what it's signing. In February, The Register reported a similar case: a platform-featured app exposed 18,697 records including 4,538 student accounts. Ten months later, the same pattern scaled to five digits.
Lovable's response framed this as an "industry-wide configuration issue—our systems were not breached, apps were never hacked." Every word is true. The unanswered question: when a platform's core selling point is that users don't need to understand security, what responsibility does the platform bear for the security baseline of its output?
The Same Message for Two Audiences
The report's advice for users is plain: before an app touches real data, review access permissions and user isolation table by table; review again after major changes. Automated checks cannot replace professional security assessment.
For the industry, one sentence suffices: vibe coding lowered the cost of *writing* apps by an order of magnitude, but the cost of *configuring* them properly did not drop—the gap between the two is these 24,664 open doors.
Practical tip: if you've used such a platform to build an app with a database, check one thing today—open your app's Supabase console and see whether RLS is enabled. It costs nothing and takes twenty minutes. The result determines whether you're part of that 24,664.