Home › Blog › Research · Incidents

Vibe Coding Risk Is Documented: What the Record Shows (2025–2026)

Security problems in AI-built apps are no longer a prediction. Between 2025 and 2026, researchers documented exposed databases, leaked keys and broken access control in apps built with AI coding tools. The newest finding, published by UpGuard on September 25, 2026, counted 16,326 Supabase databases with tables readable by anyone.

The newest finding: 16,326 readable databases

UpGuard gathered about 300,000 domains showing signs of Supabase usage and identified 16,326 databases exposing readable tables. Reading them required no login and no exploit. Over half of those databases had indicators of personal data, and smaller shares included passwords or authentication tokens.

UpGuard's explanation centers on defaults. Tables created programmatically through the API, which is how coding agents interact with Supabase, do not enable Row Level Security (RLS) by default, while tables created in the Supabase Table Editor now do. It also found public API keys shared in client-side code and used as if they were secret keys. UpGuard's reading is that these sites are created by AI coding agents while the humans behind them are unaware of the configuration, although it did not state what share of the 16,326 came from each cause.

Supabase's view, as reported by TechCrunch on the same day, is that its projects are secure by default and that security is a shared responsibility: customers control how their own projects are configured.

Lovable and CVE-2025-48757: inadequate RLS in 170 apps

In March 2025, security researcher Matt Palmer scanned 1,645 Lovable projects and found that 170 of them, about 10.3%, had inadequate Row Level Security settings, across 303 endpoints. Depending on the app, the exposed data could include personal information, API keys and access tokens for third-party services, and financial details such as transactions and subscriptions, including the ability to modify payment status. The CVE (CVE-2025-48757) was published on May 29, 2025.

In the researcher's words, the root cause is the failure to enforce or maintain secure default RLS configurations for user projects.

5,600 public apps scanned: secrets and personal data in the open

In October 2025, Escape.tech analyzed more than 5,600 publicly available apps built on platforms including Lovable, Base44, Create.xyz, Vibe Studio and Bolt.new. It reported more than 2,000 vulnerabilities, over 400 exposed secrets and 175 instances of personal data, including medical records, IBANs, phone numbers and emails. According to Escape, most of the critical weaknesses, such as exposed Supabase tokens, misconfigured APIs and missing row-level security, were reachable through public endpoints without authentication.

Escape is a security vendor, so these numbers are best read as one team's scan of public apps, not as an audit of every app on those platforms.

Moltbook: a database open to read and write

On January 31, 2026, Wiz Research found that Moltbook, an AI social network, had a Supabase database without Row Level Security policies and with API credentials in client-side JavaScript. The exposure included 1.5 million API authentication tokens, 35,000 email addresses and 4,060 private messages between agents, and researchers confirmed both read and write access.

Moltbook's founder had said publicly: “I didn't write a single line of code for @moltbook. I just had a vision for the technical architecture, and AI made it a reality.” After Wiz disclosed the issue, it was fully fixed in about 3 hours 12 minutes, across four patches.

Lovable's API flaw: 48 days between report and disclosure

A separate case, reported by The Next Web, involved a Broken Object-Level Authorization (BOLA) flaw in Lovable's API. Any free account could read other users' profiles, public projects, source code containing hardcoded Supabase database credentials, and AI chat histories, for projects created before November 2025. The flaw was reported on March 3, 2026 through Lovable's bug bounty program and stayed unpatched for 48 days until public disclosure in April 2026. This is a different issue from CVE-2025-48757.

Lovable first said it had not suffered a data breach and described the exposed data as intentional behaviour, then pointed to documentation about what public means and to the bug-bounty triage. Later the same day its management issued a partial apology.

Veracode: working code is not secure code

Veracode's 2026 GenAI Code Security Report, published on July 28, 2026, found an average security pass rate across models of 56%, barely changed from 55% in its first report. Roughly 44% of AI code generation tasks introduced a risky security vulnerability in tests. Modern models now generate syntactically correct code nearly 100% of the time, but secure coding is not following the same curve.

The 2025 report, which evaluated over 100 large language models, had already found that 45% of code samples failed security tests and that security performance remained flat regardless of model size or training sophistication.

What these cases have in common

In the app cases above, the apps were live with real users' data and the same few gaps repeat: Row Level Security missing or incomplete, keys meant to be public used as if they were protection, and authorization checks that were never written or never tested from the outside. None of these gaps stops an app from working in a demo, which is why they survive until someone outside the team tests them.

• Turn on Row Level Security for every table, then write a policy for each operation.
• Keep service-role and other secret keys out of browser code and repositories.
• Test your own app as a logged-out visitor and as a second user before launch.
• Add a security review step to every launch checklist, whoever or whatever wrote the code.