Skip to content
Security & Trust

UpGuard found 16,326 Supabase databases with tables outsiders can read

UpGuard suspects AI coding agents built many of the apps but cannot prove it. Accountability stays with the business.

W
WebPulse Newsroom
AI-assisted · 4 min read
Share on X LinkedIn
UpGuard found 16,326 Supabase databases with tables outsiders can read

AI-generated image for WebPulse. About our images

In brief
  • UpGuard Research probed about 300,000 domains showing signs of Supabase use and found 16,326 databases with readable tables. Over half showed signs of personal information, judged from table layouts.
  • UpGuard suspects AI coding agents built many of these apps, but its scans do not prove this and it did not measure how many. It calls this a working theory.
  • Ask which apps use a hosted database, who checked their access rules, and whether a person signs off before AI-built apps touch customer data. The business stays accountable.

UpGuard argues that many of these exposed sites were built by AI coding agents, and that the people who run them do not understand how their database is set up. Its scans do not prove that. Even so, the idea deserves attention from anyone who owns customer data. If an agent picks the settings and no one reviews them, the company still holds the risk.

What UpGuard found

UpGuard Research studied Supabase, a hosted database service popular with people who build apps using AI coding tools. The team collected about 300,000 unique domains that showed signs of using Supabase. It then asked each one for a table called "users".

Across that set, the researchers identified 16,326 databases with readable tables. By WebPulse's own arithmetic, that is about 5% of the domains probed, or roughly one in eighteen. It is not a rate for all Supabase sites, only for this set.

The researchers judged the contents from table layouts rather than reading every row. Over half showed signs of personal information. A smaller share held passwords or login tokens. A very small number looked like they held credit card data.

The people behind the tables

UpGuard checked a handful of cases to confirm the data was real. It notified owners where the exposure was significant. The examples show who bears the cost.

A valet service in the US Northeast exposed more than 100,000 customers. Records included phone numbers, license plates and visit history. A Canadian immigration coaching service exposed nearly 5,000 user records, and 884 of them included a plaintext password.

One database belonged to a consulate run by an African national government. It held records for 25,000 people, described as a vulnerable population. One field showed which emergency housing location each person was staying in.

The people in these tables had no say in how their data was stored.

Why UpGuard points to AI agents

UpGuard says whether a site served consumers or businesses made no difference to the kinds of data exposed. Some industries did differ. Ecommerce and restaurants were most likely to expose personal information and payment integrations. Unlicensed online betting was more likely to leak passwords and credentials.

The point UpGuard draws is that the security settings do not track the business. Owners know their own line of work but not how their database is configured. It names AI coding agents as the link across the sites.

That is an interpretation. The researchers also said their scans do not show that every affected site was built by an AI agent. They did not measure how many were. Treat it as the researchers' working theory, not a measured share.

The mechanism they describe is specific. Supabase now switches on row-level security, the control that limits who can read each row, for tables made in its web editor. UpGuard says tables created through the programming interface do not get it by default. That is the route coding agents use. Even with the control on, it must be set up correctly. UpGuard also lists public keys used as if they were secret among the failure modes.

An old pattern, possibly a new decision-maker

UpGuard compares this to Amazon S3 and GitHub. S3 grew partly through defaults that made reading and writing data easy. GitHub grew with a public-by-default model. Both led to years of exposure. The pattern is the same: defaults chosen for convenience, then leaks at scale.

What may be new, in UpGuard's telling, is who accepts the defaults. It may be an AI agent, not a person. The decision moves to software, but the accountability does not.

UpGuard says Supabase's valuation reached $10 billion by June 2026, helped by its popularity with AI-assisted builders. Popularity is not the fault. It means one small setting can be repeated across many organisations.

Limits of the study

Four limits matter. UpGuard assessed contents from table layouts, so the counts show potential exposure. It probed for a "users" table, which skews results toward personal data. It worked from public web signals, so it covers only sites with visible Supabase traces. And it did not establish how many sites were built by AI agents.

What to ask your team

Start with inventory. Ask which apps in your company, or built by contractors and marketing teams, use a hosted database. Include prototypes that became real.

Then ask who checked the access rules on each one. Ask whether row-level security is on, and whether any secret keys sit in public web code.

Finally, ask whether your policy requires a human sign-off before an AI-built app touches customer data. If an agent wrote the app, a person must still own the settings.

The lesson is short. Speed can be handed to a machine. Answering for the data cannot.

Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: UpGuard.