UpGuard found it 16,326 Supabase databases with readable tables, on approximately 300,000 domains that showed signs of use of the platform. The researchers queried each domain looking for a table called users and, from the schemas of the tables reached, they obtained the types of data present: in over half indications of personal data appear in the databases, and to a lesser extent also passwords and authentication tokens. There are very few cases with plausible credit card data, while more often the schemes indicate an integrated payment system, non-sensitive data in itself that indicates where an attacker could insert themselves into financial transactions.
The causes indicated are policy of row-level security (RLS) absent or ineffective, the control that establishes who can read the rows of a table, and the improper use of public keys. UpGuard connects the phenomenon to coding agents: in the Supabase report it is the database most recommended by Claude Code, and the growth of the platform on this front multiplies the instances that expose data. “Security settings are the same regardless of the type of business,” UpGuard writes, “because people, who know what type of business they advertise, don’t understand the configuration of their database.” UpGuard points out, however, that the scan does not establish that every site involved was built with a coding agent.
Supabase database misconfigurations have been known since 2025. In March of that year, developer Matt Turner reported misconfigured databases in apps built with vibe coding platform Lovable (CVE-2025-48757), and in February 2026 Wiz discovered that the database behind Moltbook, the social network for AI agents, left 35,000 email addresses and 1.5 million API authentication tokens readable. Supabase, an open source platform built on PostgreSQL, reached a valuation of $10 billion in June 2026.
How the databases were found
Previous investigations have worked on small samples or within single vibe coding platforms. Modern Pentest examined 107 Y Combinator startups and found personal data exposed in 28% of cases, Symbiotic Security analyzed 1,072 apps built with vibe coding and found 39 with tables readable via the Supabase public key, and Escape found 175 databases with personal data across approximately 1,400 apps. Starting with Lovable or Replit, UpGuard claims, leaves out sites created with Claude Code and OpenAI’s Codex, and for this reason the investigation looked for standalone sites on their main domain, which include real businesses with real customers.
A Supabase site can be recognized by the key names and database addresses in public JavaScript files, but verifying this for every site on the Internet requires significant resources. UpGuard therefore relied on data already collected by others: BuiltWith, which recognizes the technologies used by sites, and the Chrome UX Report on BigQuery, from which it obtained the raw JavaScript files in which to search for keys and addresses. About 300,000 unique domains emerged, each of which was asked for the users table, chosen because almost every application has users. The database responds in one of three ways: no data accessible, due to a security policy or because it is no longer operational; a results page; or the indication that users does not exist, with the name of another accessible table as a hint. With over 16,000 databases listed, the content analysis was based on table schemas rather than reading each row.
What data is exposed
Ecommerce and catering are the sectors that most often expose personal data and have integrated payment systems, in line with what these activities must deal with. Unlicensed online betting tends to leave more passwords and other credentials readable, and less intuitive is the high risk for industrial goods, i.e. the sites most likely to sell to corporate customers and handle their data. The type of offering, B2B or B2C, does not change the categories of data exposed, and the geographical distribution is global: the severity varies, with better practices in Europe, where data protection laws tend to favor them, and more exposures in developing regions. The sample also leans towards personal data because the search started from the users table, UpGuard admits.
To ensure the tables contained real data, UpGuard examined some where the metadata suggested relevant exposures, and where it confirmed them, alerted application owners. Among the verified cases there is an Indian site similar to OnlyFans: the users table counts 65,467 peoplewith name, email, date of birth and address, as well as fields for driving licence, passport, PAN card number and last four digits of Aadhaar. Linked to this data are payment accounts (PayPal, Payoneer, Zelle, crypto wallets, bank) and Stripe account information, including the last four digits of the cards, while a second table, messages, contains over 100,000 private messages to and from creators of adult content.
A Philippines-based OTP service exposes more than 2,000 users, with emails, phone numbers and wallet balances, and more than 100,000 text messages, some of which appear to be private communications between people with no connection to the service. UpGuard had already described the cybercrime chain, of which SIM farms are a link: many SIMs receive OTP codes to register fake accounts and inflate users with scams and illegal activities on Telegram and other social networks. Also according to UpGuard, a valet parking service in the United States exposes more than 100,000 customer records with contacts, license plates and visit history, a Canadian immigration service has nearly 5,000 user records with 884 plaintext passwords, and an African government consulate has data on 25,000 people, including addresses and emergency accommodation locations.
Supabase has since made RLS active by default on tables created from the Table Editor interface. Tables created programmatically via API, the channel with which coding agents interact, do not activate it automatically, and even with RLS active, protection depends on correctly configured policies and credentials. UpGuard believes the cases are likely to continue, given that the typical vibe coding user has little software engineering experience and interacts with the agent more than the code. Those who use Supabase can verify their exposure with the platform’s security documentation, in particular the advisors and the API security guide.
