UpGuard Study Finds 16,326 Supabase Databases Exposing Readable Tables

UpGuard on September 25, 2026 published research identifying 16,326 databases hosted on the Supabase platform that expose readable tables to the public web, in what the security firm describes as the largest study of its kind.

The study was written by Greg Pollock, UpGuard’s Director of Research and Insights. It reports that more than half of the exposed databases showed indicators of some personally identifiable information, a smaller percentage included passwords or authentication tokens, and a very small number contained plausible credit card data. More commonly, researchers found indicators that a payment system was used by a site, which the report notes was not sensitive in itself but flagged cases where attackers could insert themselves into financial transactions.

Supabase offers hosted Postgres databases as a service and, according to the report, has become a core component of web applications built through AI-assisted “vibe coding.” UpGuard states that Supabase reached a $10 billion valuation as of June 2026 and characterizes it as the database product most recommended by Claude Code. In cases where it determined a significant exposure, UpGuard said it notified the application owners.

Research Methodology

Rather than cataloging apps bearing the watermarks of vibe-coding platforms such as Lovable and Replit, the researchers sought standalone websites operating on their own primary domains. The report states this approach was intended to capture sites built with AI coding agents such as Claude Code and OpenAI’s Codex, as well as sites operating as real businesses that collect customer data.

Sites using Supabase can be fingerprinted by looking for Supabase key names and database addresses in public JavaScript files, according to the report. The two data sources that proved most valuable were BuiltWith, which performs technographic fingerprinting across the web, and the Chrome UX Report dataset on BigQuery, which provides raw JavaScript files that can be queried for Supabase indicators such as database addresses and API keys.

Across those two methods, the team gathered roughly 300,000 unique domains with indicators of Supabase usage and queried each for a “users” table, reasoning that “users” is a common table name. Each database returned one of three responses: no accessible data; no “users” table, but a hint naming an accessible table; or a page of results. Given the size of the candidate set, the researchers used table schemas to assess the types of data potentially exposed rather than reading every row, and scraped front-page content with an AI model classifying each business model as B2B, B2C, hybrid, or N/A.

Documented Data Exposures

To validate that the exposed schemas contained real data, the researchers investigated a subset of databases where metadata indicated a likelihood of meaningful exposure. The report details several cases.

One exposed database was tied to a one-time-passcode service in the Philippines linked to a SIM farm, an operation the report describes as part of the cybercrime supply chain in which large volumes of phone SIMs are used to register fake accounts on various platforms. The database contained over 2,000 users with email addresses, phone numbers and wallet balances, plus over 100,000 SMS messages containing one-time passcodes, sender IDs and SIM codes.

According to the report, 95% of the messages were passcodes, while between 2,000 and 2,400 sampled messages were real person-to-person texts, nearly all ridesharing communications between drivers and passengers in the Philippines. The report states those people appear unrelated to the SIM farm and were instead exposed because of the huge volume of phone numbers passing through it.

Another database belonged to a valet service based in the US Northeast that used Supabase as its CRM back end. It exposed over 100,000 customers, each with a phone number, about 43,000 of them with email addresses and full names, and about 78,000 with license plate numbers, along with each customer’s visit history, lifetime value, tip history and a free-text notes field. A staff table held 665 employee records with email addresses, phone numbers and push tokens. About 11% of the exposed email addresses (4,560) were on third-party corporate domains, including identifiable regional employers such as universities and Fortune 500 companies, the report states.

A third misconfigured database contained 25,000 users with personal information and physical addresses for a consulate operated by an African national government, including a field identifying the emergency housing location where individuals currently reside. UpGuard notes that the nature of the operation means the exposed people belong to a vulnerable population.

A fourth database, for a service providing relocation and immigration coaching for people moving to Canada, held nearly 5,000 user records. Almost all contained full names, email addresses, phone numbers and dates of birth, and 884 records also stored a plain-text password, according to the report.

Prior Research and Exposure Patterns

In March 2025, developer Matt Turner reported widespread misconfiguration of Supabase databases created by the vibe-coding platform Lovable, an issue tracked as CVE-2025-48757. Since then, Supabase has enabled row level security by default for tables created in its Table Editor interface, the report states. Tables created programmatically through the API, which is how coding agents interact with Supabase, do not enable the protection by default, and row level security must also be configured and use credentials correctly to protect data.

The study cites several earlier, smaller examinations of the same exposure vector. Modern Pentest scanned 107 Y Combinator startups and found 28% were exposing personal information. Symbiotic Security scanned 1,072 vibe-coded apps and found 39 with tables readable using the Supabase public key. Escape found 175 databases leaking personal information across roughly 1,400 vibe-coded apps. Red Access scanned 380,000 URLs and found 5,000 accessible apps, 2,000 of which were exposing sensitive corporate, operational, or personal data. In February 2026, Wiz discovered that the Supabase database behind Moltbook, a social media site for AI agents, was leaking 35,000 email addresses and 1.5 million API authentication tokens.

These misconfigurations are happening worldwide rather than concentrating in one region, according to the report, though severity is not uniform: Europe’s data protection laws tend to drive better practices, while developing regions tend to have more data leaks. By industry, ecommerce and restaurant sites were the most likely to handle personal information and integrated payment systems, while unlicensed online betting sites were more likely to leak passwords and other credentials. Whether a site advertised a consumer or business service made no difference in the kinds of data likely to be exposed, the report states.

UpGuard concludes that data leaks are the multiplicative product of a technology’s ease of misconfiguration and the size of its user base, and that the impacts land across the world, across every industry and across every business model when humans or their AI coding agents misunderstand how to implement sound security configuration.