Grok Build, the AI coding agent developed by Elon Musk-led SpaceXSI, has exposed confidential business data and personal records belonging to thousands of users as well as SpaceXSI itself, Ars Technica reported. Security analysis conducted by Claroc revealed that 94.7 percent of applications created and deployed through the platform were set to public access by default. Although Grok Build intended to protect projects by generating randomized web addresses that only direct link holders could open, structural flaws in the URL design enabled attackers to sequentially guess and enumerate thousands of active project addresses without requiring any authentication.

The exposed data went well beyond plain code snippets, spanning sensitive medical examination results, business proposals, galley checklists from a cruise operator, and employee names along with identification numbers. The exposure also compromised 24 internal applications built by SpaceXSI itself. Among the leaked corporate records were confidential onboarding documents for enterprise sales staff, an internal territory map detailing 594 client accounts, advertising proposals, and internal system addresses utilized by sales personnel.

This incident highlights significant disparities between Grok Build and competing coding assistants, such as Claude Code from Anthropic and Codex from OpenAI, which enforce strict access controls, default-private policies, and restrictions against transmitting complete version control repositories off-premises. The revelation follows an earlier incident in July 2026, when security researcher cereblab discovered that Grok Build version 0.2.93 command-line tools uploaded entire 12-gigabyte repositories, including commit histories and configuration files, totaling 5.1 gibibytes of unrequested data into an xAI cloud storage bucket. The repeated vulnerabilities emphasize that organizations adopting autonomous AI development tools must verify privilege isolation and access governance before deployment.