Grok Build Left 35,578 Apps Open, and the Default Was the Leak

No one broke in. A sharing option chosen in advance on Grok's publish panel, combined with app addresses that live in public DNS, turned "anyone with the link" into something close to anyone at all.

Somewhere on grok.me, an ICU patient briefing sat on a page that anyone could open without signing in. It was one of 35,578 apps built with Grok Build that security firm Claroc found open to the public in research published on November 10. That was 94.7% of the 37,571 Grok apps still online when Claroc tried them. Most were harmless games, calculators and websites. Claroc puts the number holding private data at roughly 650 to 800, and 24 of the open apps were built by staff at SpaceXSI, the company behind Grok, and at X.
The tempting headline is that an AI coding tool leaked data. The more accurate one is less dramatic and more useful: a default setting did it. Nothing in Claroc's report describes a vulnerability being exploited to get at these apps. They were published, with an access level that the product had proposed, at addresses that turned out to be far easier to find than their owners assumed.
How a private link becomes a public one
Grok Build lets someone describe an app in a chat and get a working version back. SpaceXSI (formerly xAI) opened it to every plan, on the web and on mobile, in August. Each published app gets its own grok.me address, and the publisher chooses who can see it.
In the paid account Claroc used, clicking Publish on an app Grok built proposed "Anyone with the link" as the access setting. "Invited users" and "Only you" were available, but the widest option was the one already selected.
Link sharing is a security model with a single assumption: the link stays secret. It works reasonably well for a document link pasted into a private chat. It works badly when the secret part of the link is published to the world, and that is where Grok's address design matters. Grok puts the random string on the left of the domain, as a subdomain. Subdomains are DNS names, so every visit triggers a DNS lookup, and passive DNS services collect those lookups from the resolvers they run or buy data from, then make them searchable. That is how Claroc assembled a list and tried 43,551 Grok app addresses.
Claroc contrasts this with Google, which places the random string on the right side of the domain, in the path. Paths do not travel in DNS queries, so there is nothing for a passive DNS database to record. The two products both say "anyone with the link." Only one of them keeps the link reasonably private once people start using it.
What was behind the doors
The SpaceXSI and X apps included an onboarding plan marked confidential and a sales territory map of 594 accounts and their pipeline. There were also X advertising proposals, one marked confidential and addressed to a US federal agency. For a company whose own product created the exposure, it is an awkward list.
Beyond SpaceXSI and X, Claroc identified 30 organizations whose internal data appeared in open apps; 26 of them have more than 1,000 employees and six are in the Fortune 500. The examples are the kind of thing people build when a tool makes building cheap: a hospital staff roster, the ICU briefing, a fire department shift board. One contractor's timesheet app carried every employee's account in its code with the password in plain text, admins included, 18 accounts in all.
Exposure was not limited to reading. In apps Claroc tested, an outside visitor could add a row to a contact list that had no sign-in. Another customer list did have a "Continue with Google" button, but any Google account could read, edit and delete its records, because the generated code added every signed-in user to the team before checking whether they belonged there. A login screen, in other words, is not the same thing as access control, and an AI that writes the login screen will not necessarily write the check behind it.
SpaceXSI's defense, and what it reveals
When a reporter asked about the research on November 10, SpaceXSI said Grok apps stay private until their creator chooses to share them, that publishing takes several steps, and that employee accounts default to "Only you." The company took down the 24 staff apps. Other users' apps were left as they were.
Claroc's reply is one sentence long and hard to argue with: "Extra clicks on a path that ends with access already set to everyone do not protect the data." A confirmation dialog protects against accidental publishing. It does nothing for a user who meant to share with a colleague and did not realize the address could be found by strangers.
The detail about employee accounts is the most telling part of the response. If SpaceXSI's own staff start from "Only you," the company has already decided that the narrow default is the right one for data it cares about. The paid customers in Claroc's test started from the widest one. SpaceXSI has not said whether it will change that.
Second time, same shape
This is the second time this year that a Grok Build default has decided where user data ends up. In July, The Register reported a researcher's finding that the Grok Build command-line tool sent file contents without redaction to a Google Cloud Storage bucket used by SpaceXAI, and Elon Musk promised the uploaded data would be purged. The July problem sent data to the company. The November problem exposed it to everyone. But the shape is the same: the safe state was the one users had to reach for themselves.
That matters more as AI tools stop being developer toys. Agents and app builders are now reaching people with no reason to think about DNS or row-level permissions, and security around AI agents has become a market in its own right. The Grok Build case is a reminder that a product does not need a breach to produce a breach-sized result. A default chosen for convenience, applied tens of thousands of times, gets there on its own.
The limits of the count
Claroc's numbers deserve some care. The 650-to-800 estimate comes from a model checked against a random sample the team read by hand, not from reviewing every app. The vast majority of open apps contained nothing sensitive, and an open calculator is not a leak. The findings also come from one firm; Ars Technica reported on them, but no second team has published its own scan.
None of that changes the core point, which does not depend on the exact count. Claroc's suggested fixes are modest: serve shared apps on hard-to-guess paths, offer an option limited to a company's email domain, make "Only you" the default, and warn when an app has a database but no sign-in. For organizations, the practical step is to find out which AI app builders employees use, search DNS data for apps carrying the company's name, and test from outside whether they open. Building an internal tool has become a matter of an afternoon. Checking who can see it still takes a deliberate look.
Sources
Comments
Post a Comment
Please read our comment policy before commenting.