UpGuard finds roughly 16,000 Supabase databases exposing personal data online

UpGuard finds roughly 16,000 exposed Supabase databases

Security research firm UpGuard has discovered around 16,000 databases hosted on the development platform Supabase that exposed personal information to the public internet, TechCrunch reports. Supabase, which lets web and app developers store and run their databases, has become a popular backend for AI-generated applications, and the findings point to a systemic gap between how easily those apps can be built and how well they end up secured.

Names, passwords and license plates left open

According to UpGuard, the publicly accessible data included names, addresses, phone numbers and user passwords, with a smaller number of authentication tokens also surfacing. The affected databases spanned a striking range of use cases. Among those identified by the researchers were an Indian adult streaming site that exposed private conversations between users and sex workers, a US valet service whose records covered thousands of license plates, and an immigration and relocation service that exposed the contact details of people who had used it.

One exposed database belonged to an African government's consulate in France, UpGuard said. Another was tied to a virtual SIM farm used to intercept text messages carrying one-time passcodes for online account verification — infrastructure typically associated with scam and phishing campaigns.

UpGuard said the majority of the exposed datasets appear to be located in the United States, but characterised the problem as worldwide. The research builds on earlier findings that documented exposed Supabase-hosted databases belonging to Y Combinator startups and other popular apps.

A vibe-coding security problem

Supabase reached a $10 billion valuation earlier this year, growth TechCrunch attributes in large part to developers hosting AI-coded, or "vibe-coded", apps on the platform. The company has drawn criticism before over how it handles user security, with widely documented cases of customers misconfiguring databases or unknowingly exposing them to the broader internet — in some instances amounting to millions of records each.

The pattern echoes a long history of breaches traced to improperly configured storage servers, databases and websites, which have previously leaked sensitive military emails, immigration and visa applications, classified government files, hundreds of thousands of driver's license scans and children's personal information. What has changed, TechCrunch argues, is scale and accessibility: AI tools let people with little security knowledge ship working apps quickly, but the generated code can carry flaws, and the surrounding services often require configuration the developer may not know exists.

Supabase: secure by default, shared responsibility

Supabase's chief information security officer, Bil Harmer, said the company had not yet seen UpGuard's research but insisted its projects are "secure by default." He framed security as a shared responsibility between the platform and its customers. "We provide secure defaults and tooling, and customers control how their own projects are configured," Harmer said, adding that Supabase notifies affected customers when security issues are discovered.

"Security at Supabase is never finished. We care deeply about getting it right, and we'll keep making it easier for every developer to ship securely," he said. UpGuard researcher Greg Pollock said the research was important for raising awareness of data exposures on the platform. Supabase has made changes over the years, including strengthening platform and user access controls on databases.

Why it matters

The finding illustrates where security is failing in the AI coding era: not primarily inside the model-generated code, but in the infrastructure decisions wrapped around it. Backend-as-a-service platforms such as Supabase remove enormous amounts of plumbing work, but they still hand developers real access-control decisions, and a single misunderstood setting can silently publish an entire user table. With roughly 16,000 affected databases on one platform alone, misconfiguration is no longer an edge case — it is a predictable outcome of onboarding non-specialist builders at scale. That will likely increase pressure on platforms to proactively detect and warn about exposed data rather than relying on customers to configure things correctly, and it is a concrete warning to anyone shipping an AI-built app that handles other people's information.