Software · AI-built apps

She built an app with AI. Its database was open to anyone.

A Hamburg security firm checked what apps built with Lovable hand to a stranger who never logs in. From 159,041 projects it found records with email addresses readable on 7,403 sites — among them passport details, payment data and 200,000 chat messages. The study was published on November 2; this week, the internet finally read it.

THE BUYER'S PULSESoftware7 min read
A woman wearing headphones works on a laptop at a bright home desk
The new app builder: one person, one laptop, one AI tool. The study found that thousands of apps built this way went live with the database rules that decide who can read their data switched off or set wrong.

Photo: Kaboompics, via Pexels (Pexels License — free to use). The woman pictured is not one of the builders in the study; the image shows the kind of solo, at-home app building the report examines.

Start with the person in the photograph, because she is the protagonist of this story even though she appears in no report. She is not a software engineer. She has an idea — a booking page for her studio, a small shop, a tool for her team — and an AI service that turns plain-language descriptions into working applications. She types what she wants, the service builds it, and by the weekend her app is live, with real users and a real database behind it. What she almost certainly never saw is the one setting that decides who is allowed to read that database. A study published in Germany this month measured how often that setting was left open — and the answer is: thousands of times.

The study comes from Crissy Field GmbH, an application-security firm in Hamburg, and was reported with the German business newspaper Handelsblatt on November 2, 2026. The researchers took the open web archive Common Crawl and identified 159,041 projects built with Lovable — the Swedish AI platform, founded in Stockholm in 2023, that generates web applications from text prompts — published between December 2025 and September 2026. Then they asked a narrow, practical question of each one: what does this app hand to a total stranger who has not logged in?

159,041Lovable projects identified in Common Crawl (December 2025 – September 2026)
34,545Of those projects used Supabase as their database
24,664Sent database records the moment the page loaded, with no login
7,403Sites where read-only requests returned records containing email addresses
10,560Tables holding those email records
30%Of the closely examined sites — a figure the researchers state is NOT a confirmed leak rate

What the researchers actually did

The method had two steps, and both used only what any visitor's browser already receives. First, the team recorded what each site sent when its page loaded. Of the 159,041 projects, 34,545 used Supabase, the hosted database service most common among Lovable apps — and 24,664 of those sites, about seven in ten of the Supabase group, returned database records on page load, before any login or click. Some of that data is meant to be public — a product list is not a secret — so the researchers went one step further: with additional read-only requests, using the same public key each page itself ships to the browser, they looked for records containing email addresses, the simplest marker of personal data.

That second step is where the figure at the center of the story comes from. Records containing email addresses turned up in 10,560 tables across 7,403 of the 24,664 closely examined sites — 30 percent of them. The researchers did not break into anything: every request was read-only, made with the key the app itself gives away. They also did not download the datasets in full, only parts of them, and they do not name the affected apps or publish how the data can be reached — so the counts they report are a floor for what was readable, not a total.

Passports, payment data and 200,000 chat messages

The examples the report describes are not contact forms. A German lifestyle forum exposed more than 80,000 user profiles — names, email addresses, dates of birth — together with more than 200,000 chat messages. In one case, a single app held some 280,000 customer email addresses. And in what the researchers call a particularly serious case, a service provider exposed more than 1,000 records containing identity-document details: passport numbers, expiration dates and nationalities. Payment data also appears among the categories the team found readable without a login.

Data of that kind is the raw material of phishing, ticket fraud and identity theft, as the researchers themselves note. What they could not establish is equally part of the record: there is no evidence in the report that any of this data was actually misused. The doors were open; whether anyone walked through them, beyond the researchers' own read-only requests, is unknown.

The caveat is part of the finding

Crissy Field is unusually explicit about what its headline figure does not mean, and the qualification belongs in the body of the story, not in a footnote. The 30 percent is not a confirmed leak rate. Some of the email addresses counted may be public contact addresses or test data. The method also misses personal records that contain no email address, so the same figure can understate the exposure in other places. And the Common Crawl sample is not representative of all Lovable apps: it covers the projects the archive happened to capture in that ten-month window. What the study establishes is narrower and, for that reason, harder to dismiss: in a sample of 159,041 projects, thousands of live apps would hand personal records to anyone who asked with the app's own public key.

One missing rule: Row Level Security

No Lovable system was hacked, and no Supabase system was hacked — the report alleges no breach of either company. The mechanism is more ordinary, and more instructive. An app built this way talks to its Supabase database directly from the visitor's browser, using a public «anon» key that is embedded in the page on purpose: the key is meant to be public. What decides who can read what is a feature inside the database called Row Level Security — RLS — a set of rules, table by table, that says which rows each kind of visitor may see. Where RLS is disabled or badly configured, the public key is enough: anyone holding it can request the whole table. Supabase's own documentation is blunt about this, instructing developers to enable RLS on every table in an exposed schema.

Responsibility is where the accounts diverge. Lovable told Handelsblatt that the problem is common across the industry, and under the platform's terms the customer is responsible for their app and how it handles data. In the Hacker News discussion that carried the study around the world this week, Lovable's head of security acknowledged the mechanism directly: the root cause in these apps was insufficient Row Level Security — the very thing the platform's pre-publish security scan is meant to catch, with a free deeper scan also available. The scan existed; in thousands of published apps, the rule was still missing.

This exact failure was already reported in 2025

None of this is the first warning. In March 2025, researchers scanning 1,645 sites from Lovable's own showcase found 170 whose tables were open to anyone; the issue was assigned a vulnerability identifier, CVE-2025-48757. Crissy Field's sample begins eight months after that report — December 2025 — which is the uncomfortable part of the timeline: whatever checking the ecosystem added in between did not reach the long tail of people publishing apps. The builders changed faster than the safety net did.

Why a November 2 study is news again today

The chronology deserves honesty, because the dates are the story's second lesson. Handelsblatt published the findings in German on November 2, with the firm's own report — carrying a more recent analysis and updated figures — alongside it. The story then travelled the way technical findings now travel: resubmitted to Hacker News on November 14, it reached the front page and, by one account of the thread, passed 750 points with nearly 300 comments arguing over where the blame sits — the platform that makes building effortless, the database whose safety depends on one setting, or the builders who never knew the setting existed. The summaries and explainers of the last two days, this one included, are the echo of that argument.

The check takes minutes, because the exposure is simple

The same simplicity that made the study possible makes the individual check fast — a point the technical explainers of the last two days keep repeating. An app's owner can request one of their own tables from the database API using nothing but the public key the page already ships: if rows come back that a stranger should not see, Row Level Security is off or a policy is wrong. Supabase's own Security Advisor flags tables with RLS disabled. And the researchers' scope carries its own warning in reverse: the study only tested what a visitor gets without an account. What a second, signed-up user can read inside the same apps — one customer reading another customer's rows — is a separate question the report does not answer, and the same missing rules decide it.

Strip the argument down to what is verified and a simple picture remains. AI tools have reduced the distance between «I have an idea» and «my app is on the internet» to an afternoon. The database behind the app did not get any simpler in the same afternoon: it still runs on access rules written for a world of professional developers, and one missing rule is enough to publish the customer list with the product. The study's most durable number may not be 7,403 sites or 30 percent. It may be the 24,664 apps that answered a stranger before he had even logged in — and the fact that finding them took a web archive, a public key and patience, nothing more.

𝐆𝐄𝐓 𝐅𝐑𝐄𝐄 𝐅𝐔𝐋𝐋 𝐌𝐀𝐆𝐀𝐙𝐈𝐍𝐄 𝐀𝐓 thebuyerspulse.com

Sources

  • Crissy Field GmbH — «Lovable: Many Apps Built on the Platform Expose Sensitive Data», research report, November 2026 (updated figures).
  • Handelsblatt — Lina Knees, report on the Crissy Field research, November 2, 2026 (in German).
  • Supabase documentation — Row Level Security (RLS).
  • Coverage and explainers of the report: Frontier Now (November 15, 2026); Keybrum; dev.to summaries (November 15–16, 2026).