Grok Build coding agent leaks oodles of confidential info

Security researchers at Claroc found that Grok Build, SpaceXSI’s AI coding agent, exposed more than 35,000 user-created apps and their stored data to anyone who found their links. The links were meant to be difficult to guess, but their randomized words appear before the grok.me domain. That placement lets passive DNS databases record and reveal the addresses; putting the secret portion after the domain, as Google and several other services do, would keep it out of those records. Claroc identified 43,601 Grok Build links, nearly 95 percent of which opened apps publicly. Its findings conflict with a SpaceXSI representative’s claim that apps are private by default: Claroc’s tests found that users had to change the default “Anyone with the link” setting themselves.

Exposed material included medical examination records for thousands of workers, employee names and ID numbers, sales prospects and financial notes, customer proposals, transport orders and margins, and sensitive defense-industry procedures. Twenty-four affected apps belonged to SpaceXSI itself, including confidential sales plans, account maps, internal tool addresses, and advertising proposals. The company removed those apps after receiving questions about the exposure, while Claroc said it was contacting other affected organizations. Similar link-sharing risks may exist in other AI app-building services, though Claroc found especially sensitive information on Grok Build. Users should restrict access to invited people unless they intend an app to be broadly available.