Grok Build Flaw Left 35,000 AI-Made Apps Open, Including SpaceX Data
Researchers say about 35,000 apps made with Grok Build, Elon Musk's AI app builder, could be opened by anyone, exposing business, medical and SpaceX data.
Tens of thousands of apps built with Grok Build, the AI app-making tool from Elon Musk’s artificial intelligence business, could be opened by anyone who went looking for them, security researchers say. The exposed data reportedly included confidential sales material from the company itself, along with medical test results and employee records belonging to outside organizations.
Ars Technica reported on Tuesday that a researcher had found roughly 35,000 Grok Build apps “wide open” to anyone. The findings come from Claroc, a cybersecurity firm that called its analysis “GhostShares,” according to Kantan News and the news aggregator Zetik, which both summarized the research.
What happened
Grok Build lets people describe an app, game, website or dashboard in a chat with Grok, the company’s AI assistant, and have a working version built for them. In an August 19 announcement, xAI said the feature was available on every plan, on the web and on mobile. Each published app gets its own address on the grok.me domain. Creators choose who can open it: “just you,” “anyone with the link” or “the whole internet.”
According to Kantan News, the main protection for link-shared apps was obscurity. Each address was built from a random combination of words that, in theory, only people who had the link would know. The problem was where that identifier sat. Because it was placed in the subdomain, the part to the left of grok.me, rather than in the path after it, every app’s address showed up in passive DNS records. These are public logs of internet domain lookups that anyone can search, which meant the supposedly hidden addresses could be found with standard discovery techniques.
Kantan News reported these figures from Claroc’s scan:
- 43,601 grok.me addresses identified
- 37,571 active apps scanned
- 35,578 of those served their data without asking for a login
- 1,993 had restricted access settings
Ars Technica’s summary put the number of open apps at about 35,000, in line with those figures.
What was exposed
Among the open apps were 24 internal tools built by employees of the company that runs Grok, which Ars Technica referred to as SpaceXSI, the name used for the merged SpaceX and xAI business. According to Kantan News and Zetik, those apps held a confidential onboarding plan for a new member of the enterprise sales team, a sales territory map covering 594 accounts with revenue projections, addresses of internal management tools and several advertising proposals. Kantan News said one dashboard tracked a $1.6 million sales target.
The exposure went well beyond Musk’s companies. Kantan News reported that Claroc identified records linked to six Fortune 500 companies, the U.S. Navy, healthcare providers and a large cruise line operator. The material reportedly included medical examination results for 3,784 patients, more than 4,000 vehicle shipping and logistics records, cruise ship kitchen checklists, and employees’ full names and personnel numbers. The outlets did not name the outside organizations, and their accounts could not be independently verified.
How the company responded
Zetik reported that Claroc first notified the company on a Sunday and received no reply. After journalists began asking questions, the company took down the 24 links to its own internal apps. Claroc said nearly all of the other exposed data was still reachable when its report was published, according to Zetik. The company did not appear to have issued a public statement on the findings in the coverage reviewed.
Why it matters
This is the second privacy problem for Grok Build this year. In July, an independent researcher found that the command-line version of the tool, a separate coding assistant for programmers, was uploading entire code repositories to company cloud storage. Musk said at the time that the uploaded data would be deleted, and the company later released that tool’s source code.
The new findings point to a broader risk as AI tools let people with no programming background build and publish software in minutes. Many such users may not know that a link shared with “anyone” can end up in public records, or that a random-looking address is not the same as a password. Security experts have long warned that keeping a link secret is a weak way to protect sensitive data, and the Grok Build case suggests that tools built for mass audiences may need stricter defaults.
