A security researcher's free Lovable account became a skeleton key. With five standard API calls, they accessed source code, database credentials, and AI chat histories from projects belonging to Nvidia, Microsoft, and others. The platform, a $6.6 billion 'vibe-coding' startup, confirmed no breach. The fault lay in system architecture.

The issue was a Broken Object Level Authorization flaw. Lovable's APIs, particularly for listing and fetching project contents, did not validate user ownership against the requested resource. Public project visibility, initially intended to share final applications like GitHub repos, inadvertently exposed the entire development workspace—including chat histories where developers routinely paste secrets for AI context.

Lovable's engineering timeline reveals the trade-offs. Early design prioritized collaboration and 'sparking ideas,' making projects public by default. Enterprise controls and a shift to private-by-default arrived later. A critical misstep occurred in February 2026: a backend permissions unification accidentally re-exposed chats on historical public projects. Internal HackerOne reports were initially misclassified, delaying a fix for 48 days after the first alert.

This incident underscores a structural tension in AI-assisted development platforms. The engineering imperative for rapid iteration and simple abstractions—like a unified 'public' flag—can conflict with granular security postures. While Lovable has patched the specific API endpoints, the episode serves as a case study for data engineers: default permissions and object-level authorization are not afterthoughts, but foundational design decisions with material consequences.

Source: Webpronews