Sentrail
Research database

AI Coding Security Incidents

Sourced records that distinguish confirmed real-world incidents from product vulnerabilities, research findings, and unverified reports. Every record shows its evidence class, impact boundary, citations, remediation, and update history.

Confirmed real-world incidentConfirmed product vulnerabilityResearch findingUnverified report

Published records

Evidence before claims.

4 verified, sourced records

Confirmed real-world incidentResolvedLovable · 2026-02-03

Lovable public-project access regression

Lovable confirmed that a backend regression made chat history and source code for public projects potentially accessible to an authenticated Lovable user who had the project link.

access controlexposed source code
Review evidence
Confirmed real-world incidentRemediatedReplit · 2025-07-18

Replit Agent production database deletion

During a public build exercise, Replit Agent deleted a production database despite an active code freeze. The affected builder documented the event and recovery; Replit subsequently described product controls intended to separate development and production databases.

destructive agent actionagent/tool permission issue
Review evidence
Confirmed product vulnerabilityDisputedLovable · 2025-04-15

CVE-2025-48757: disputed Lovable RLS-policy vulnerability

The NVD record for CVE-2025-48757 describes insufficient row-level security in generated Lovable applications. The record also preserves Lovable's dispute that application builders, rather than the platform, own generated-app data protection.

RLS misconfigurationauthorization failure
Review evidence

Taxonomy

Tags describe the affected security control. They do not change the evidence classification.

RLS misconfigurationaccess controlagent/tool permission issueauthentication failureauthorization failuredestructive agent actionexposed source code