Safety#Site building#AI generation

Study finds 24,664 Lovable apps serving database records without a login

German security firm Crissy Field scanned 159,041 Lovable projects and found 24,664 apps handing out database records to anyone, passport numbers included.

A peeling pink Lovable poster on the wall of a derelict building, torn paper exposing the steel frame behind it

German application-security firm Crissy Field examined 159,041 projects built on Lovable, the Swedish AI app builder, and found 24,664 Supabase-backed apps returning database records as soon as the page loaded, before a visitor logged in or clicked anything. The readable data included more than 1,000 identity-document records at one service provider — passport numbers, expiry dates, nationalities — plus more than 80,000 profiles and over 200,000 chat messages from a German lifestyle forum. The researchers stress they sent only read-only requests with the public anon key that every frontend already ships to the browser: the report does not allege a breach of either company, nor any exploit chain.

Key points

  • The sample comes from Common Crawl: 159,041 Lovable projects online between December 2025 and September 2026. Of those, 34,545 used Supabase, and 24,664 returned database records on load — about 71% of the Supabase group, by the report’s own figures.
  • Email addresses were the easiest marker: records containing them turned up in 10,560 tables across 7,403 sites, 30% of the sites examined more closely. The researcher states plainly that this is not a leak rate.
  • The story reached 750 points and 298 comments on Hacker News after it was resubmitted on November 14 (as of 06:30 UTC on November 15), with the argument turning on blame: Lovable, Supabase, or builders with no security background at all.

Background

Lovable is a Stockholm company founded in 2023, grown out of the open-source GPT Engineer project, renamed and launched publicly in late 2024; it turns plain-language descriptions into web apps with user accounts and databases. Supabase is its most common database service, according to the researchers: 34,545 of the 159,041 projects in the sample used it. Cybernews reports that Lovable has delivered more than 50 million projects and adds 1.2 million a week.

The publication order matters. Crissy Field ran the study with the German business daily Handelsblatt, which published the German story on November 2; the report page carries newer figures and says so in a note. The English version took off after it was resubmitted to Hacker News on November 14 and hit the front page that afternoon.

This is not Lovable’s first data-exposure story. In February 2026, The Register reported that an exam-and-grades app listed on Lovable’s Discover page exposed 18,697 user records; the company’s response then was that projects get a free pre-publish security scan, but users decide whether to act on the findings. On November 10, Lovable’s own blog introduced pre-publish scans, deep scans and one-click fixes — we noted it in our November 11 digest. That was a product release, not this: the tooling gives owners a way to find and fix findings, while the report looked at projects that have been live for a year or more.

For indie developers, Supabase as the default backend is a normal cost-saving choice; we list it among the free tiers in our low-cost indie stack for 2026. The catch is that the same default path turns one configuration mistake into a batch problem.

The facts

Crissy Field’s report gives these headline numbers; the counts reflect what the researchers actually downloaded, so the underlying datasets may hold more.

MetricFigureScope
Projects identified159,041Common Crawl, December 2025 to September 2026
Using Supabase34,545About 21.7% of the sample
Records returned on load24,664About 71% of Supabase projects, no login or interaction
Records with emails7,403 sites, 10,560 tables30% of the 24,664 sites examined in detail
Identity documentsMore than 1,000 recordsOne service provider; passport numbers, expiry, nationality
Forum data80,000+ profiles, 200,000+ messagesGerman lifestyle forum; names, emails, birth dates, gender
Single-case emails280,000Customer emails readable at one provider

The method ran in two steps: record what a site sent on first load, then send further read-only requests with the same public anon key to see what else came back. The researchers say they did not download full datasets, only parts of them, and they withhold which apps are affected and how to reach them, because publishing that would make further access easier. The scope was Supabase only; the database features Lovable introduced in late 2025 were not examined, though the authors expect similar misconfigurations could exist there.

The limitations are stated just as plainly. Emails are a useful marker for records that may hold personal data, but public contact addresses and test data get counted too, while personal records without an email get missed; how far those two effects cancel out is unknown, the authors say. A Common Crawl sample is not representative of all Lovable apps either. So the 30% applies only to the 24,664 sites inspected further and should not be read as an overall leak rate.

What others say

Lovable’s position is that its AI is not the problem. According to Cybernews and Trending Topics, the company describes the finding as a known, industry-wide Supabase configuration issue and points to its Security Scan and AI security reviewer. Neither outlet carries a verbatim company quote, and as of November 15 we found no public statement about the report on Lovable’s own site. The primary report says only that Handelsblatt contacted Lovable with the findings.

Lovable’s documentation is worth reading alongside that. Its Security overview page, viewed on November 15, lists built-in Quick scan and Deep scan, row-level security (RLS) checks for database access, and a project security view. Whether those tools exist and whether an existing project actually has RLS turned on are different questions.

The Hacker News thread spreads the blame around. Commenter SOLAR_FIELDS argues Supabase shares responsibility and links the org’s 2021 discussion “Option to disable anon key access to database,” which was still receiving replies on October 21, 2026. hn_throwaway_99 sets a harder test: these defects need no zero-day chain, just sensitive data in a database anyone online can read, so only real liability changes behavior. Oras adds the other half — AI-built apps are numerous, cheap to make and growing daily, attackers need no sophistication, and users have no idea what they are using. berkes asks the procedural question: has any of this been reported to regulators under GDPR?

A second dataset comes from UpGuard. BleepingComputer reported on September 28 that the risk-management firm scanned roughly 300,000 domains showing signs of Supabase and found more than 16,000 databases readable from outside, over half with indications of personal data. UpGuard’s own summary: the common thread is that these sites are created by AI coding agents and the humans are unaware of the configuration.

Our take

The framing of this Lovable exposure matters, because it decides what readers do. This is publicly accessible data observed by researchers, not a major Lovable breach: no exploit chain, no stolen credentials, and the anon key ships to every browser by design. The numbers need the same care — 200,000-plus is a message count, not users; 1,000-plus is a record count, not passport holders; 24,664 counts apps that returned records on load, and some of that content was meant to be public.

Three groups should act. Anyone who built something on Lovable or Supabase should check table by table whether RLS is on, then try reading the database from a private window with the anon key in the frontend; anything you should not see is a problem. Teams holding customer data in these apps should stop treating a passed pre-publish scan as an audit, since scans cover known patterns rather than your business logic. Agencies shipping these apps should put an access-review step in the handover, and repeat it after significant changes.

What interests me is the leverage that sameness creates. Hand-built projects differ, so one working probe hits one site; template defaults let the same technique run against thousands of apps at near-zero marginal cost. That is the same asymmetry we wrote about when AI cut the cost of finding flaws without cutting the cost of fixing them.

Several things remain open: how many of the 24,664 apps have been fixed, how much data was actually read, whether anyone has filed under GDPR, and whether Lovable will check existing projects rather than wait for owners to act. The report does not name affected apps, which is defensible but leaves the people in those databases unable to check for themselves.