Supabase Data Exposures Show the Security Cost of AI Vibe Coding

Introduction

AI “vibe coding” has made it possible for people with limited software experience to launch websites and applications quickly. But an application that works is not necessarily an application that protects its users. Security firm UpGuard says it found around 16,000 databases hosted on Supabase where some degree of personal information was exposed. In some cases, the data was accessible from the public internet and included names, addresses, phone numbers, as well as a smaller number of passwords and authentication tokens.

The findings do not indicate that Supabase itself was breached. Based on the available material, the central issue was project configuration, including database permissions and access policies. Even so, the research highlights the practical difficulty of maintaining a shared-responsibility security model as more inexperienced developers deploy AI-assisted applications.

Key points

• A broad set of exposed projects. UpGuard said the databases covered multiple countries and use cases. Although most of the datasets appeared to be located in the United States, the issue is global rather than confined to one market.
• Sensitive information was involved. The research found names, addresses and phone numbers, along with a smaller number of passwords and authentication tokens. The examples included private conversations on an Indian adult-streaming service, license-plate records from a U.S. valet company, and contacts linked to an immigration and relocation service.
• Different organizations and services were affected. UpGuard also identified a database used by an African government consulate in France and another linked to a virtual SIM operation that intercepted text messages used for online account verification. The material does not establish that these exposures led to additional attacks.
• AI can amplify configuration mistakes. Generated code may produce a functional interface without correctly implementing row-level security, authentication, least-privilege access or data separation. Developers who are new to backend systems may also mistake a convenient default for a secure default.
• Security remains a shared responsibility. Supabase Chief Information Security Officer Bil Harmer said the platform provides secure defaults and tooling, while customers control how their projects are configured. He also said the company notifies affected customers when issues are found.

Why it matters

The larger lesson is not that AI-generated software is inherently unsafe. Rather, AI is changing who can ship software and how quickly they can do it, while the complexity of database security remains largely unchanged. A public endpoint, an overly permissive policy or an access rule copied from an example can leave data exposed long after an application goes live.

Developers using AI-generated code should manually verify anonymous access, least-privilege rules, password hashing, token handling and the separation between test and production data. Platforms can also help with stronger defaults, configuration scanning, clearer warnings and timely notifications. If the speed of vibe coding is not matched by security checks, traditional configuration errors may become more common and harder to detect at scale.

Source: TechCrunch AI