What did Claroc find?
Grok can build a working app from a description in the chat; every published app gets its own subdomain. According to Claroc, it tried 43,551 such app addresses. 37,571 were still online, and 35,578 of those (94.7%) opened without sign-in. Only 1,993 answered with an HTTP 401 error, which an app set to Only you or Invited users returns.
According to Claroc, most open apps were games, landing pages and calculators. About one in ten, however, held sensitive, internal or personal data. Among them were 24 apps built by staff at SpaceXSI, the Grok maker formerly called xAI, and at X; SpaceXSI took them down after the report. Claroc also found internal data from 30 further identifiable organizations. The research names none of them, and neither do we.
Why were so many apps open?
Claroc cites two reasons. First, the default: in the paid account the researchers used, the publishing step offered Anyone with the link, Invited users and Only you, with Anyone with the link already selected. Second, an app address does not stay secret: opening it triggers a DNS lookup of the subdomain, passive DNS services record such lookups, and they can be searched by pattern.
According to Claroc, SpaceXSI told a reporter that Grok apps are private until their creator shares one; publishing takes several steps, and employee accounts are set to Only you by default. Claroc counters that at the end of this path, access for everyone was preselected in the account it tested. Claroc also describes how AI apps drift with further prompts: a harmless app can hold customer data the next day.
Does a sign-in protect the app automatically?
Not necessarily. Claroc built two test apps of its own with made-up data. A contact list got a database but no sign-in; from a private browser window, the researchers added a row that the next visitor saw. A customer list for a four-person sales team got sign-in buttons from Grok, including Continue with Google. The generated code, however, added every signed-in person to the team before checking membership.
An outside Google account that had never been invited could thus see every customer, change one record and delete another. Claroc stresses that it ran these write and delete tests only on its own apps. The research also found a timesheet app whose script file, downloaded by every browser, contained accounts with plain-text passwords. A sign-in button therefore says nothing about whether the authorization check behind it is correct.
How do companies check their Grok apps?
Claroc recommends several checks, which The Hacker News summarizes. They are simple and can be started without special tools. According to Claroc, an app set to Only you returned a 401 error to a stranger for both the page and its data function, and unpublishing took an app offline within seconds.
- Find out which AI app builders staff use; many need neither IT setup nor a company account.
- Set every Grok app holding internal data to Only you or Invited users.
- Test from a private browser window and with an outside account: a stranger must see an error page.
- Search loaded script files for names, email addresses and passwords; no credentials in client-side code.
- Unpublish apps that nobody owns.
What does this mean for companies in Germany and the EU?
The open apps described by Claroc included staff lists, shift boards with names and badge numbers, recruiting lists with personal email addresses and a patient briefing. When a company's staff load such data into a publicly reachable app, personal data is affected. That is a matter for the data protection officer and the internal procedure for reporting data protection incidents.
Whether a notification obligation applies in a given case cannot be said in general and belongs in an expert assessment; this article is not legal advice. It makes sense to secure affected apps immediately, document the incident and involve data protection. As a rule, staff and customer data should not go into AI-built apps without a reviewed sign-in and authorization check.
Which rules belong in the AI policy?
Claroc describes the core problem like this: an app is made in a chat window and published from the same window, often in minutes, by people whose job is a sales territory or a duty roster. IT sees none of it. An AI policy should therefore explicitly cover AI app builders, not just chat assistants.
In our view, a few clear rules work well: which builders are approved, that every published app has an owner, that apps holding internal or personal data go live only with restricted sharing and a reviewed sign-in, that no passwords or keys sit in the code, and that apps are reviewed regularly for sharing and content. If you are unsure which apps already exist, start with an inventory.
Keep it verifiable





