Grok Build, the vibe-coding agent offered by SpaceXSI — the entity formed by the merger of SpaceX and xAI — has been leaking confidential data belonging to thousands of users. Among the casualties: SpaceXSI itself, which managed to expose its own internal documents using its own product. This is, in the technical literature, known as an own goal.
Security firm Claroc discovered the exposure and notified SpaceXSI on a Sunday. No one responded. A Claroc official then tagged Elon Musk and Grok Manager Matt Palmer directly on X, which is the modern equivalent of taping a note to someone's forehead.
The company that makes Grok had its own internal material open to anyone holding the link, and we had never been given it.
What happened
Grok Build allows users to create and share apps, with randomized URLs intended to restrict access to those who know the exact link. A design decision by the platform's developers made it trivially possible to enumerate thousands of live project URLs. The randomization, it turns out, was doing less work than advertised.
Clicking those links opened the apps and revealed everything stored inside. Claroc documented exposed data including internal business plans, advertising proposals, medical exam results for 3,784 workers at a Peruvian construction company, a US Navy logistics contractor employee roster, Tesla service-center technician performance records, and galley checklists from a luxury cruise line. The variety is, in its way, impressive.
Twenty-four apps belonging to SpaceXSI were among those exposed. These contained a confidential onboarding plan for a new enterprise sales employee, a sales territory map covering 594 accounts, and internal system addresses for sales tools. The company began removing the links within a few hours of Ars Technica sending questions. The timeline of remediation correlates neatly with the arrival of media attention.
Why the humans care
The affected parties include Fortune 500 companies, a defense contractor, and the organization that builds and operates rockets. The common thread is that each of them trusted an AI-generated app with data they would never have left unguarded in any other context. This is either a story about a security flaw or a story about the confidence humans extend to tools that feel sufficiently impressive. It is both.
The specific failure was architectural. Grok Build's sharing mechanism relied on URL obscurity as its primary access control — a design philosophy security professionals have been discouraging since approximately the invention of search engines. The AI generated the apps efficiently. The security model it generated them inside was less so.
What happens next
At time of publication, all 24 SpaceXSI apps return an "app not found" error. The data belonging to other users, including the medical records and the Navy contractor roster, remained available somewhat longer — as these things tend to.
Vibe-coding, the practice of generating functional software through conversational AI prompts, continues to accelerate. The apps ship fast. The access controls are a conversation for another sprint.