A security researcher found 16 vulnerabilities — including inverted access-control logic — in a Lovable-generated EdTech app, exposing 18,697 users' records and letting anyone delete accounts or grade exams

1. Instruction given

   Not a single developer's prompt to an agent — the flaw lived in the AI-generated backend of an exam-question and grading app that Lovable itself featured on its public Discover showcase, built through Lovable's standard app-generation flow on a Supabase backend.

2. Expected behavior

   Restrict administrative functions and other users' records to authorized, authenticated accounts; an unauthenticated visitor should have no access to protected data or destructive actions.

3. Actual behavior

   Security researcher Taimur Khan found that a remote procedure call function meant to block non-admin users from sensitive functions instead did the opposite: it blocked every logged-in user while letting unauthenticated requests through. Khan said the same inverted-logic error recurred across multiple critical functions in the app, which had no Supabase row-level security or role-based access control implemented at all. The app had logged over 100,000 views and about 400 upvotes on Lovable's own Discover page before Khan's disclosure.

4. Damage

   Khan found 16 vulnerabilities in total, six rated critical, exposing 18,697 user records — 14,928 of them with unique email addresses, including 4,538 student accounts and 10,505 enterprise users, with 870 users' full PII exposed. Per Khan's findings, an unauthenticated attacker could read every user record, send bulk emails through the platform, delete user accounts, grade students' exam submissions, and reach organizations' admin emails. Affected users reportedly included people from UC Berkeley and UC Davis, and possibly K-12 minors. Khan says he first reported the issue through Lovable's support, where the ticket was closed without response; Lovable told The Register it received "a proper disclosure report" on the evening of February 26, 2026 and acted "within minutes." Lovable CISO Igor Andriushchenko said every project includes a free pre-publish security scan but that acting on its recommendations is "at the discretion of the user," and noted the project also contained code Lovable did not generate and a database Lovable does not host. Lovable said it had contacted the app's creator, who was "now addressing the issue" as of publication; no independent confirmation of a fix is reported, and whether anyone besides Khan accessed the exposed data is not established.

Classification

Agent

Lovable

Failure mode

Security Vulnerability

Root cause

Logic Error

Domain

Backend

MIT AI Risk Repository (our mapping)

2.2 AI system security vulnerabilities and attacks · 2.1 Compromise of privacy by leaking or correctly inferring sensitive information taxonomy

Source

Security Research

StupidLLM. "STUPID-2026-0151: A security researcher found 16 vulnerabilities — including inverted access-control logic — in a Lovable-generated EdTech app, exposing 18,697 users' records and letting anyone delete accounts or grade exams." AI Agent Incident Database, 2026. https://www.stupidllm.com/incident/STUPID-2026-0151/ (accessed 2026-10-31). CC BY 4.0.

← STUPID-2026-0150By ID