Lovable Apps Left the Door Open on 7,403 Sites, and Nobody Built the Lock
The biggest agent-adjacent story on Hacker News this Saturday was a database permission default. Crissy Field, a German security research firm, took 159,041 Lovable projects from Common Crawl (Dec 2025 to Sep 2026) and looked at the 34,545 that used Supabase. 24,664 of those sent database records the moment the page loaded. With further read-only requests, the researchers found tables containing email addresses on 7,403 of those sites, 10,560 tables in total, about 30%. The English report hit #1 on HN on Nov 14 with 487 points. The underlying research is not new: Handelsblatt first reported it in German on Nov 2, and the English report with updated figures is dated Nov 3. Report: https://www.crissyfield.de/research/lovable-en/
The examples are bad. A German lifestyle forum exposed more than 80,000 profiles and more than 200,000 chat messages. Event organizers and restaurants exposed tens of thousands of ticket and food-sale records. A property manager exposed AI-generated summaries of internal emails about tenants. A service provider exposed more than 1,000 records with passport numbers. The authors say the 30% is not a confirmed leak rate, since email addresses can be public contacts or test data, and that a Common Crawl sample does not represent every Lovable app. They also withheld access details on purpose. No response from Lovable appears in the report.
This isn't a Lovable exploit. The apps did what they were built to do, and nobody set the rules for who gets to read which row. Lovable's marketing says you don't need technical knowledge, while its terms put responsibility for apps and data on the customer, and that gap is the whole story. A founder who can't read code can't audit row-level security either. The agent generated a working app and a permission model in the same pass, and only the first one got tested, because only the first one shows up on screen.
For every app-building agent, the fix is a verifier for the part users never see. Before an app touches real data, something has to check what an anonymous visitor can read, the same way the build already checks that the page renders. Until platforms ship that check by default, the advice in the report is the only advice: review the access permissions before you put real people's data in.
← Back to all articles
The examples are bad. A German lifestyle forum exposed more than 80,000 profiles and more than 200,000 chat messages. Event organizers and restaurants exposed tens of thousands of ticket and food-sale records. A property manager exposed AI-generated summaries of internal emails about tenants. A service provider exposed more than 1,000 records with passport numbers. The authors say the 30% is not a confirmed leak rate, since email addresses can be public contacts or test data, and that a Common Crawl sample does not represent every Lovable app. They also withheld access details on purpose. No response from Lovable appears in the report.
This isn't a Lovable exploit. The apps did what they were built to do, and nobody set the rules for who gets to read which row. Lovable's marketing says you don't need technical knowledge, while its terms put responsibility for apps and data on the customer, and that gap is the whole story. A founder who can't read code can't audit row-level security either. The agent generated a working app and a permission model in the same pass, and only the first one got tested, because only the first one shows up on screen.
For every app-building agent, the fix is a verifier for the part users never see. Before an app touches real data, something has to check what an anonymous visitor can read, the same way the build already checks that the page renders. Until platforms ship that check by default, the advice in the report is the only advice: review the access permissions before you put real people's data in.
Comments