Supabase Exposure Study Raises Security Questions for AI-Built Apps
Security researchers say they found around 16,000 Supabase-hosted databases with some personal information exposed to the public web. TechCrunch reports the finding from UpGuard, which examined data accessible through the development platform. The number describes databases in the researchers’ survey; it is not a count of individual victims, confirmed intrusions or databases created with AI tools.
The AI connection lies in how quickly people can now assemble and deploy apps. Supabase has attracted developers building so-called vibe-coded projects, and TechCrunch frames the exposure as a warning about configurations that can be overlooked in that workflow. An AI-generated app may have vulnerable code or require security settings its builder does not understand. The report does not establish that AI caused each of the exposures UpGuard found.
What the researchers say they found
UpGuard told TechCrunch that the exposed records included names, addresses, phone numbers and, in a smaller number of cases, passwords or authentication tokens. The examples span very different activities: private conversations on an Indian adult-streaming site, license-plate data from a US valet service and contact information from an immigration and relocation service. One database reportedly belonged to an African government’s consulate in France. Another was associated with a virtual SIM operation used to intercept verification texts.
Those examples make the finding more concrete than a platform-wide count alone, but they do not describe the contents of every database in the survey. UpGuard said most exposed datasets appeared to be in the United States while the problem extended internationally. The source does not provide an audited total of affected people or say that all exposed records were downloaded by third parties. Public accessibility is itself a serious condition; whether information was exploited is a separate question.
TechCrunch describes a history of developers misconfiguring Supabase projects and making data public without intending to do so. It also notes that other researchers had previously found exposed datasets on the platform. The new contribution is UpGuard’s broader measurement and its selection of examples, not the first discovery that database configuration can go wrong.
A platform dispute over responsibility
Supabase’s chief information security officer, Bil Harmer, told TechCrunch the company had not seen UpGuard’s research when contacted. He said projects are secure by default and framed protection as a shared responsibility: Supabase provides defaults and tools, while customers control how their projects are configured. He also said the company notifies affected customers when issues are discovered. Those are the company’s statements, not an independent resolution of which party is responsible for each database.
The distinction between a default and a deployed configuration matters. Even if a new project begins with protective settings, a builder can change access rules, connect other services or release an application without adequately testing its data paths. Conversely, the presence of exposed data on a hosted service does not automatically prove a defect in the service’s underlying infrastructure. TechCrunch’s account does not provide enough project-level detail to assign fault across the roughly 16,000 cases.
The company has made platform changes over time intended to strengthen access to databases, according to the report. UpGuard researcher Greg Pollock said the research was meant to raise awareness of the exposure problem. That stated goal does not substitute for disclosure to every affected person, and the source does not describe a complete remediation schedule for all of the datasets it identified.
The AI-development link needs careful reading
The report ties the issue to rapid app development aided by AI, but the data presented is about Supabase-hosted databases. It does not break out how many exposed projects were vibe-coded, which model or tool was involved, or whether errors arose from generated code, manual configuration or both. Treating every record as evidence of an AI-specific failure would go beyond the source.
What the story does show is a security pressure point where fast application creation meets live personal data. The examples range from contact details to potentially sensitive conversations and credentials, making the consequences of an exposed database dependent on the project rather than on a single platform-wide headline. UpGuard’s findings and Supabase’s response leave a practical question for each affected operator: what information was reachable under its deployed settings, and who could access it? TechCrunch does not yet supply those answers for individual projects.
