Grok Build leaked 35,000 apps, including SpaceX's own secrets
A security flaw in Grok Build, the vibe-coding agent from SpaceX's xAI, left tens of thousands of applications built by its users open to anyone on the internet, including applications built by SpaceX staff that exposed the company's own confidential business plans.
Security firm Claroc discovered the problem and documented it in a report titled "GhostShares 2: Grok Apps Exposed xAI's Business Secrets and People's Private Data", published on 10 November 2026. The findings were reported the same day by Ars Technica, and covered by Engadget on 11 November.
The mechanism of the exposure
Every application built with Grok Build receives its own hostname, in the form <random-words>.grok.me. The random words that were supposed to keep an application private sit directly in the hostname itself.
That matters because hostnames are visible in passive DNS records, the logs of which domains have been looked up. Anyone who can read those records can collect the hostnames, and with them, the addresses of applications that were never meant to be found.
Other sharing systems do not work this way. Google Docs, Claude artifacts and ChatGPT shared conversations keep the secret part of a link in the path, after the domain name, so it never appears in DNS logs. Grok Build put it in the hostname, where it was logged every time the application was visited.
The scale of the exposure
Claroc collected 43,601 grok.me application names from passive DNS data. It tried to open 43,551 of them. Of those, 37,571 were still online, and 35,578 of them, or 94.7%, opened with no sign-in at all. Only 1,993, or 5.3%, returned an HTTP 401 error, meaning they asked for authentication.
In other words, the overwhelming majority of Grok Build applications that were still live could be opened by anyone who knew the address, and the addresses were sitting in public DNS data.
Claroc estimates that around 650 to 800 of the exposed applications, roughly 2% of those it checked, hold private data, which it classifies as high or critical risk.
SpaceX's own applications were exposed
Among the exposed applications were 24 built by staff of SpaceX's xAI itself. Claroc says those applications contained a confidential enterprise-sales onboarding plan, a sales territory map covering 594 accounts, eight X advertising proposals, one of them addressed to a U.S. federal agency and marked confidential, four live advertiser boards, one of which tracked a $1.6 million sales goal, seven campaign reports, and two X brand-safety playbooks.
The exposure was not limited to one company's files. Other applications that Claroc found open included, as reported by Ars Technica, a Tesla service-centre scorecard naming 22 technicians; 3,784 worker medical exams from a Peruvian construction company; a logistics roster from a Navy contractor; a hospital staff app with a searchable roster of more than 3,500 employees; an ICU patient briefing; and a fire-department shift board.
Disclosure and response
Claroc reported the problem to SpaceX's xAI through its security.txt contact. It reported the enterprise-sales onboarding plan and the federal-agency advertising proposal on 8 November, and the other 22 staff applications on 9 November.
SpaceX's xAI never replied directly to Claroc. It did answer a reporter's questions on 10 November, claiming that applications are "private by default", that publishing an application takes several steps, and that employee accounts default to "Only you".
The 24 SpaceX's xAI applications were still live on the evening of 9 November. They were pulled within hours of Ars Technica emailing questions about them, and by the time of publication all 24 returned "app not found". Other users' applications remain open.
Implications
The incident shows how a small design choice can undo a whole privacy model. Grok Build's applications may have been described as private by default, but their addresses were being recorded in passive DNS the moment anyone visited them, and the vast majority of the applications that were still live asked for no sign-in at all.
For the people affected, the consequences are concrete. The exposed applications included employee medical records, a hospital staff roster of more than 3,500 people, an ICU patient briefing, and a Navy contractor's logistics roster, information that was never meant to be public.
For SpaceX and X, the leak is especially awkward: the company's own staff built applications that exposed its confidential sales plans, its advertising proposals, including one to a U.S. federal agency, and its brand-safety playbooks, all reachable from the open internet.
Outlook
The 24 SpaceX's xAI staff applications are now offline, returning "app not found". But the underlying problem reaches far beyond them. Tens of thousands of other users' applications remain open, and Claroc estimates hundreds of them hold private data.
The fix, in principle, is straightforward: secrets belong in the path of a link, not in the hostname. Until Grok Build changes how its share links work, every application built with it risks the same exposure, discoverable through passive DNS, and open to anyone who looks.
More in AI

Musk and Sweeney tell Nvidia's board: save Jensen

Microsoft's Lazarus: 7 edited neurons turn an AI agent into a sleeper backdoor
