Skip to content
Artwork for Deep Dive
Deep Dive · Yesterday · 29 min

Vibe Coding Security: How AI-Built Apps Leak Your Data

Open a website someone built with an AI app builder this year. Type your email, a birthday, a document number. In November 2026 a security firm working with Handelsblatt went looking for apps where the lock was never put on. It scanned 159,041 apps built with one tool, Lovable. 24,664 of them handed database records to anyone who loaded the page. 7,403 held email addresses, and inside them: roughly 80,000 user profiles, 200,000 chat messages, about 1,000 identity-document records. Not one of those apps was hacked. Accessible is not the same as taken, and the researchers never established misuse. We explain the mechanism in plain English: the key sitting in every page was always meant to be public, and the lock, row-level security, is a separate step whose default depends on how the table was made. Then the five findings in twenty months that nobody had put in a row, from March 2025 to this one. Then Lovable in its own words, in date order: "pretty much guaranteed to be secure," "the 99 percent who lacked the technical skills," "at the discretion of the user," and, after the scan, "the actual blame falls on us" — set beside terms of service that cap its liability at twelve months of fees. Then Supabase, whose chief executive says no one visits the dashboard where the warning lives. Then the law: under European privacy law the person who hit publish is the controller, and California prices a breach at $107 to $799 per consumer. The strongest case against this episode gets four parts, including Firebase in 2018 and Lovable's answer that apps built today are much harder to get wrong. Then the precedents — MongoDB, Amazon's storage buckets, Firebase — and the pattern in every one: the vendor changes the default, and only the next app is safer. A ten-minute check for anyone who built one of these apps, and two dated calls. Disclosure: this show is made with Claude, and Lovable builds on Claude models. The researchers kept every case anonymous, and so do we. As of mid-November 2026. RELATED EPISODES How AI Coding Tools Leak Your Passwords to GitHub How AI Agents Actually Work Why Banks Are Blocking AI Agents (and Who Pays When One Goes Rogue) CHAPTERS 00:00 The scan 01:08 A disclosure 01:59 What this episode covers 02:25 What the scan found 05:52 How an app leaks with no hacker 08:48 The alarm nobody sees 09:15 The buried default change 10:17 Five scans in twenty months 12:50 Lovable in its own words 15:35 The contract 16:56 Supabase and the law 20:09 The case against this episode 23:47 Has this happened before? 26:09 The ten-minute check 27:48 Two calls 28:38 Three sentences, eight years apart SOURCES Crissy Field, Lovable security research (Nov 2026): https://www.crissyfield.de/research/lovable-en/ Hacker News thread on the Crissy Field report: https://news.ycombinator.com/item?id=50499017 Supabase changelog, tables no longer exposed to the Data API automatically (Apr 28, 2026): https://supabase.com/changelog/45329-breaking-change-tables-not-exposed-to-data-and-graphql-api-automatically Supabase Series F post (Jun 2026): https://supabase.com/blog/supabase-series-f Matt Palmer, statement on CVE-2025-48757 (May 2025): https://mattpalmer.io/posts/statement-on-CVE-2025-48757/ The Register, Lovable security coverage (Feb 2026) Lovable terms of service: https://lovable.dev/terms Lovable security docs: https://docs.lovable.dev/features/security Semafor, Lovable is a sitting duck for hackers (May 29, 2025): https://www.semafor.com/article/05/29/2025/the-hottest-new-vibe-coding-startup-lovable-is-a-sitting-duck-for-hackers Heuking, Vibe coding: legal issues and liability risks: https://www.heuking.de/en/news-events/newsletter-articles/detail/vibe-coding-legal-issues-and-liability-risks-of-ai-generated-software.html SecurityWeek, Appthority Firebase study (Jun 2018): https://www.securityweek.com/thousands-mobile-apps-leak-data-firebase-databases

0:00-29:40

transcript

No transcript — this publisher did not publish one.

show notes

Open a website someone built with an AI app builder this year. Type your email, a birthday, a document number. In November 2026 a security firm working with Handelsblatt went looking for apps where the lock was never put on. It scanned 159,041 apps built with one tool, Lovable. 24,664 of them handed database records to anyone who loaded the page. 7,403 held email addresses, and inside them: roughly 80,000 user profiles, 200,000 chat messages, about 1,000 identity-document records. Not one of those apps was hacked. Accessible is not the same as taken, and the researchers never established misuse.

We explain the mechanism in plain English: the key sitting in every page was always meant to be public, and the lock, row-level security, is a separate step whose default depends on how the table was made. Then the five findings in twenty months that nobody had put in a row, from March 2025 to this one. Then Lovable in its own words, in date order: "pretty much guaranteed to be secure," "the 99 percent who lacked the technical skills," "at the discretion of the user," and, after the scan, "the actual blame falls on us" — set beside terms of service that cap its liability at twelve months of fees. Then Supabase, whose chief executive says no one visits the dashboard where the warning lives. Then the law: under European privacy law the person who hit publish is the controller, and California prices a breach at $107 to $799 per consumer.

The strongest case against this episode gets four parts, including Firebase in 2018 and Lovable's answer that apps built today are much harder to get wrong. Then the precedents — MongoDB, Amazon's storage buckets, Firebase — and the pattern in every one: the vendor changes the default, and only the next app is safer. A ten-minute check for anyone who built one of these apps, and two dated calls.

Disclosure: this show is made with Claude, and Lovable builds on Claude models. The researchers kept every case anonymous, and so do we. As of mid-November 2026.

RELATED EPISODES
How AI Coding Tools Leak Your Passwords to GitHub
How AI Agents Actually Work
Why Banks Are Blocking AI Agents (and Who Pays When One Goes Rogue)

CHAPTERS
00:00 The scan
01:08 A disclosure
01:59 What this episode covers
02:25 What the scan found
05:52 How an app leaks with no hacker
08:48 The alarm nobody sees
09:15 The buried default change
10:17 Five scans in twenty months
12:50 Lovable in its own words
15:35 The contract
16:56 Supabase and the law
20:09 The case against this episode
23:47 Has this happened before?
26:09 The ten-minute check
27:48 Two calls
28:38 Three sentences, eight years apart

SOURCES
Crissy Field, Lovable security research (Nov 2026): https://www.crissyfield.de/research/lovable-en/
Hacker News thread on the Crissy Field report: https://news.ycombinator.com/item?id=50499017
Supabase changelog, tables no longer exposed to the Data API automatically (Apr 28, 2026): https://supabase.com/changelog/45329-breaking-change-tables-not-exposed-to-data-and-graphql-api-automatically
Supabase Series F post (Jun 2026): https://supabase.com/blog/supabase-series-f
Matt Palmer, statement on CVE-2025-48757 (May 2025): https://mattpalmer.io/posts/statement-on-CVE-2025-48757/
The Register, Lovable security coverage (Feb 2026)
Lovable terms of service: https://lovable.dev/terms
Lovable security docs: https://docs.lovable.dev/features/security
Semafor, Lovable is a sitting duck for hackers (May 29, 2025): https://www.semafor.com/article/05/29/2025/the-hottest-new-vibe-coding-startup-lovable-is-a-sitting-duck-for-hackers
Heuking, Vibe coding: legal issues and liability risks: https://www.heuking.de/en/news-events/newsletter-articles/detail/vibe-coding-legal-issues-and-liability-risks-of-ai-generated-software.html
SecurityWeek, Appthority Firebase study (Jun 2018): https://www.securityweek.com/thousands-mobile-apps-leak-data-firebase-databases