Grok Build Apps Exposed Sensitive Data From Thousands of Users, SpaceXSI Included

Security firm Claroc has disclosed a flaw in SpaceXSI's Grok Build vibe-coding agent that exposed sensitive data from thousands of users, SpaceXSI included. Engadget reported the findings on 11 November 2026.

The exposure came through the links Grok Build generated for the apps it produced. Those links were publicly reachable through passive DNS services, which record hostnames as they are resolved, and they appeared in publicly searchable databases. People who had never been given a link could therefore find and open the pages. Claroc calls the exposed pages GhostShares for that reason.

The leaked material ranged from employee-related records to commercial documents. Sensitive employee information came from a hospital group, a fire department and SpaceXSI itself. SpaceXSI's exposed files included a sales territory map covering 594 strategic accounts, eight X advertising proposals, four live advertiser account boards, and two copies of an X brand safety playbook along with an onboarding plan.

Access controls were thin. Only one of the exposed apps asked for a password. Grok Build created each link from four randomly generated dice words, which Claroc describes as carrying up to 38 bits of entropy, a measure of how hard a value is to guess. The words were placed to the left of the domain. That is a large search space for brute force, but it does nothing to stop a hostname from being logged by a passive DNS collector or indexed by a database crawler. Once the name is recorded, its randomness no longer matters to who can see it.

The reach extends well beyond SpaceXSI. Twenty-six companies with exposed Grok Build apps had more than 1,000 employees, and six other exposed companies appear on the Fortune 500 list. Claroc estimated that roughly two percent of the exposed material was serious enough to rate high or critical.

After Claroc tagged Musk and another company official on X, SpaceXSI took down 24 of its own apps that had been exposed through Grok Build links. Claroc stated that SpaceXSI had not taken down exposed apps left open by other Grok users and organizations.

The broader context here will be familiar to anyone who has run security reviews on link-sharing features. A capability URL, where possession of the address is treated as permission, is only as strong as the address stays secret. Addresses leak through DNS resolvers, certificate transparency logs, browser histories, referrer headers and search indexes. Long random tokens make a link harder to guess. They do not make an unauthenticated endpoint private.

In my view, the more useful question for vendors building agentic app generators is not how many bits of entropy a link carries but whether the generated application authenticates its viewers at all. A password prompt on one app out of the set Claroc examined is a narrow gate. The Grok Build case shows how quickly a casual deployment can put commercial strategy documents and personnel records on the open internet.

Still, the upside of these tools is real. Vibe coding lets people without formal software training build working internal tools, and that capability will keep spreading. The same ease of deployment is what makes default settings important. If the generator enforces authentication and treats public exposure as an explicit, reviewed choice, the productivity gain stays intact and this class of exposure becomes much harder to produce by accident.