Skip to content
AI.info

The Pulse

UpGuard Finds 16,326 Supabase Databases With Readable Tables

UpGuard says it found 16,326 Supabase databases with readable tables after testing about 300,000 domains for exposed data. Its researchers identified schema indicators of personal information in more than half of the databases and verified

UpGuard Finds 16,326 Supabase Databases With Readable Tables

AI.info Team ·

UpGuard identified 16,326 Supabase databases with readable tables, a count drawn from roughly 300,000 domains its researchers flagged as using the database platform. The company says more than half of those databases showed schema indicators of personally identifiable information, while a smaller share appeared to contain passwords or authentication tokens. The tally measures access to readable tables—not confirmed theft or misuse of every record.

How UpGuard tested 300,000 domains

Researchers used data from BuiltWith and Google’s Chrome User Experience Report to find sites with signs of Supabase use, including database addresses and API keys embedded in public JavaScript. They then queried each candidate database for a table named “users,” a common label in web applications. A response could indicate that access was blocked, that another table name was exposed as a hint, or that the database returned rows.

UpGuard assessed table schemas across the 16,326 readable databases rather than attempting to inspect every record in every table. The researchers say the sample’s personal-data skew partly reflects their decision to probe for “users” tables. They checked a smaller number of cases where metadata suggested meaningful exposure, and say they notified application owners when they confirmed significant exposures.

Sample checks found identity records and private messages

Among the examples UpGuard examined was an adult-content platform in India whose users table contained records for 65,467 people. The listed fields included names, email addresses, dates of birth, physical addresses and identity-document details, alongside payout and payment information. A separate table held more than 100,000 private messages between creators and users.

Another exposed database belonged to a US valet service. UpGuard says it contained records for more than 100,000 customers, including phone numbers, visit histories and license-plate numbers; a staff table held contact details and push tokens. The researchers also describe a Philippines-based service whose database contained more than 100,000 text messages, mostly one-time passcodes, as well as a smaller number of personal conversations.

The examples show why a database that returns readable data can create risks beyond account details. UpGuard found records connected to a government consulate, including physical addresses and emergency-housing locations, and an immigration service where 884 records reportedly contained plain-text passwords. The report does not name the organizations, and it does not establish that someone accessed or misused the data.

Configuration mistakes persist across creation methods

UpGuard traces the exposure pattern to database access controls, including row-level security policies that are missing or fail to restrict access adequately, and public keys used in ways that give them excessive permissions. Supabase has enabled row-level security by default for tables made through its Table Editor, the report says, but tables created programmatically through its API do not receive that default. AI coding agents commonly use that API, while developers may not inspect the resulting database configuration.

The findings extend earlier investigations focused on particular app-building services, because UpGuard searched for standalone sites on their own domains rather than limiting its sample to platforms such as Lovable or Replit. The report presents AI-assisted development as a common factor, but its scan does not establish that an AI tool created every exposed database or caused each misconfiguration. It also leaves open how many of the 16,326 databases remained readable on September 25, 2026, whether unauthorized parties had accessed them, and how many owners had fixed the underlying permissions.

Source

Explore

More articles