Grok Build Share Links Expose Thousands of Confidential Enterprise Applications

Security researchers at Claroc discovered that Grok Build, the AI-powered coding agent from SpaceXSI (the merged SpaceX and xAI entity), inadvertently published confidential business data through its shareable link feature. Although URLs were randomized to limit access, a platform design decision enabled enumeration of approximately 35,000 live projects, making their contents accessible to anyone who located the links.

The exposed applications included internal onboarding plans, sales territory maps covering 594 accounts, and advertising proposals from SpaceXSI itself. Beyond the vendor's own data, researchers found occupational health dashboards with 3,784 worker medical exams from a Peruvian construction firm, a Tesla service-center scorecard naming 22 technicians, a U.S. Navy ship engineering contractor roster, crew manifests from a luxury cruise line, and a hospital staff directory of 3,565 employees. A confidential hardware bill of materials for a major retail proposal and loading procedures for explosive materials from a Nordic aerospace group were also accessible.

Why it matters for GPU and AI infrastructure

The breach illustrates a systemic risk: as organizations accelerate AI agent adoption, the underlying compute and storage infrastructure often lacks enterprise-grade identity and access management. GPU clouds hosting such agents must enforce tenant isolation, audit trails, and default-deny sharing policies to prevent similar data sprawl. For buyers, this means evaluating not just model performance but the security posture of the entire inference and application stack.

SpaceXSI removed the 24 internal applications within hours of disclosure, yet the majority of third-party apps remained exposed at publication time. The episode reinforces that convenience features like one-click sharing require rigorous threat modeling before deployment on shared GPU infrastructure.