# UpGuard found 16000 Supabase databases leaking user data to the open web > Source: > Published: 2026-09-25 18:10:03+00:00 *16,000. That's how many publicly exposed Supabase databases UpGuard found leaking names, passwords, and authentication tokens, TechCrunch reported.* 16,000 is the number UpGuard came up with after scanning Supabase's public infrastructure for exposed data. TechCrunch reported on September 25 that the cybersecurity firm identified roughly that many Supabase-hosted databases publicly exposing personal data, including names, home addresses, phone numbers, passwords, and authentication tokens. Supabase is the backend platform that stores the database, handles login, and manages file storage for a huge share of apps built with AI coding tools like Lovable, Bolt, and Replit. This isn't a hack. It's thousands of developers who never turned on the lock. The examples TechCrunch cited are not abstract. UpGuard's researchers found private conversations from an Indian adult streaming site sitting in the open, thousands of license plate records from a U.S. valet parking service, contact details from an immigration and relocation firm, a database belonging to an African government's consulate in France, and infrastructure tied to a virtual SIM farm used to intercept text messages for account verification scams. Real names next to real passwords, queryable by anyone with a URL and a few minutes of curiosity. The mechanism behind almost all of it is Postgres row-level security, or RLS, a feature that decides which rows of a database table a given user is allowed to see. Supabase's own dashboard turns RLS on by default when you create a table through its Table Editor. But AI coding tools don't click through the dashboard. They write SQL migrations directly, and in that path RLS defaults to off. A founder types a prompt, the AI ships a working app in an afternoon, and nobody ever flips the switch that keeps one user from reading every other user's row. This is not a new failure mode. In May 2025, security researcher Matt Palmer scanned 1,645 live apps from Lovable's own public showcase and found that 170 of them, about one in ten, had critical RLS failures severe enough to earn a CVE, tracked as CVE-2025-48757 with a CVSS score of 8.26. One of the flagged apps was an education platform featured on Lovable's own Discover page, and it exposed roughly 18,000 users: 14,928 email addresses, 4,538 student accounts, and 870 records containing full personal information, students and educators at multiple U.S. universities among them. The exploit didn't require sophistication. The public anon key embedded in the app's own front-end code was enough to query tables directly and pull them wholesale. [Lovable's API flaw exposed private project data from the $6.6 billion AI app builder used by Nvidia and Microsoft teams](https://startupfortune.com/lovables-api-flaw-exposed-private-project-data-from-the-66-billion-ai-app-builder-used-by-nvidia-and-microsoft-teams/) Lovable, the $6.6 billion AI app builder, confirmed an API vulnerability that allowed any logged-in user to access chat histories, source code, and database credentials from projects created before November 2025. The flaw, reported via HackerOne in March 2026, exposed real-world data including information tied to a Danish nonprofit's project... - [API security vulnerability exposed Lovable project data](https://startupfortune.com/lovables-api-flaw-exposed-private-project-data-from-the-66-billion-ai-app-builder-used-by-nvidia-and-microsoft-teams/) - [Lovable AI builder data breach Nvidia Microsoft](https://startupfortune.com/lovables-api-flaw-exposed-private-project-data-from-the-66-billion-ai-app-builder-used-by-nvidia-and-microsoft-teams/) Supabase's CISO, Bil Harmer, told TechCrunch the company had not yet reviewed UpGuard's specific findings but said Supabase provides secure defaults and tooling and that security ultimately remains a shared responsibility between the platform and the customers building on it. That's technically true, and it's also the same answer every cloud provider gives after an S3 bucket left open costs someone their customer list. The dashboard default doesn't help you if you never touch the dashboard. Supabase isn't a fringe player you'd expect to stumble into this. The company closed a $500 million funding round in June 2026 that valued it at $10.5 billion, on the strength of being the default backend that AI coding assistants reach for when they need somewhere to put user data. That scale is exactly the problem. When a platform becomes the default for tools that generate an entire app's infrastructure from a text prompt, the platform's defaults become everyone's defaults, and a security setting nobody remembers to check turns into sixteen thousand open doors instead of one. Here's the thing founders building on Supabase need to hear plainly: RLS is not optional, and **Also read:** [The Fed Finally Wrote Stablecoin Rules, and Wall Street Banks Come Out Ahead](https://startupfortune.com/the-fed-finally-wrote-stablecoin-rules-and-wall-street-banks-come-out-ahead/) • [Goldman Sachs says hyperscalers will spend $1.2 trillion on AI in 2027](https://startupfortune.com/goldman-sachs-says-hyperscalers-will-spend-12-trillion-on-ai-in-2027/) • [Anthropic's founders want 50.1% voting control before their IPO](https://startupfortune.com/anthropics-founders-want-501-voting-control-before-their-ipo/) *This article is posted in [AI News](https://startupfortune.com/category/ai/), check it out for more related stories.* ## Join the discussion [Open in the community →](https://startupfortune.com/community/) Almost there. Sign in and your reply posts straight away. [Supabase Open-Sources Evals to Grade Claude Code, Codex and OpenCode](https://startupfortune.com/supabase-open-sources-evals-to-grade-claude-code-codex-and-opencode/) Supabase has open-sourced Evals, an Apache-2.0 benchmark that scores AI coding agents like Claude Code, Codex and OpenCode on real backend tasks like schema building and RLS policy fixes. Early results show Opus 5 and Kimi K3 hitting 100% unaided, while Sonnet 5 needed Supabase's context files to climb from 78% to 100%. - [how to evaluate AI coding agents](https://startupfortune.com/supabase-open-sources-evals-to-grade-claude-code-codex-and-opencode/) - [benchmarking Claude Code with Supabase](https://startupfortune.com/supabase-open-sources-evals-to-grade-claude-code-codex-and-opencode/)