Grok App
Claroc went looking for Grok-built apps and found 37,571 of them. Of that number, 35,578 were open to anyone without a sign-in.
That is not a rounding error. That is 94.7% of everything they found.
Twenty-four of those open apps were built by staff at X and at SpaceXSI, the company behind Grok. SpaceXSI has taken those 24 down. Every other user's exposed app is still up.
What Was Actually in Those Staff Apps
The researchers described three of the 24 in their report.
|
App |
What It Contained |
|
Onboarding plan |
Confidential plan for a new sales hire |
|
Sales territory map |
594 named accounts |
|
Advertising proposal |
Marked confidential, prepared for a U.S. federal agency |
That last one is the kind of document that should never have left a shared drive, let alone been published on a public subdomain.
SpaceXSI told a reporter that apps stay private until their creator shares them, that publishing takes several steps including an access setting choice and a legal disclaimer, and that employee accounts default to "Only you."
The researchers' finding stands against that. Twenty-four apps built by staff at the company and at X were open to anyone.
Beyond the Staff Apps
The exposure extends well past the two companies.
The open apps contained internal data from 30 other organizations the researchers could identify, spanning construction, energy, healthcare, and defense. Of those, 26 have more than 1,000 employees and 6 are Fortune 500 companies.
Roughly 2% of the open apps contained material serious enough to rate as high or critical.
The examples are specific and grim.
- A hospital group's searchable list of more than 3,500 employees
- An intensive care briefing on a single patient
- A fire department's live shift board
- A contractor's timesheet whose source code held every employee's password in plain text
That last one is not a misconfiguration. That is a builder generating code that embeds credentials client-side, and a user publishing it without reading it.
How the Apps Were Found
Grok can build a working app from a description typed into chat. That feature opened to every Grok plan on August 19.
Each published app gets its own subdomain of grok.me, and the creator chooses who can open it.
Here is the part that makes discovery trivial. An app's address does not stay secret. When someone opens an app, their computer looks up its subdomain in DNS. Passive DNS services record those lookups, and anyone with an account can search those records by pattern.
Grok suggests four random words as each app's name. About half the names the researchers found followed that pattern. Random names are hard to guess, but guessing is not the attack. The name is already in a searchable list.
The researchers used common subdomain discovery methods and loaded each app's first page once, without signing in.
The Default Setting Problem
In the paid account the researchers used, publishing offered three access settings.
|
Setting |
Who Can Open It |
|
Anyone with the link |
Anyone, no sign-in required |
|
Invited users |
People the owner approves |
|
Only you |
Just the creator |
"Anyone with the link" was already selected.
Contrast that with ChatGPT's app builder, which also puts each app on its own subdomain. When the researchers shared an app there, its access was set to "Only me" until they changed it. A ChatGPT app address found in DNS data leads to an app that only its owner can open.
That is the difference between a default that fails safe and a default that fails open.
What Grok's Own Protections Actually Do
To see what protection Grok adds on its own, the researchers built two test apps with made-up data and published both with "Anyone with the link" already selected.
Test One: The Contact List
The first app got a database but no sign-in. From a private browser window, without signing in, the researchers added a row. The next visitor saw it.
That is a shared database with no authentication layer at all.
Test Two: The Sales Team List
The second app, a customer list for a four-person sales team, got sign-in buttons that Grok added on its own. "Continue with Google" and "Continue with X."
But the generated code added each person who signed in to the team before checking whether they belonged. The check always passed because the membership was granted first.
An uninvited Google account, belonging to someone other than the owner, got in. It viewed every customer, edited one record, and deleted another.
That is an authorization logic flaw generated by the builder and shipped without anyone reviewing it.
The researchers ran these tests only on their own apps. They did not test whether the open apps they found would accept such changes.
The Disclosure Timeline
The researchers reported their findings to SpaceXSI on Sunday, including two of the staff apps, then reported the other 22 on Monday.
By the time they published, the company had not replied to their report. It had answered a reporter's questions about the research.
Claroc asked the company to make "Only you" the default and to add an option that limits an app to a company's own email domain. The company's response did not state that it would contact the owners of other open apps.
So Claroc started emailing affected organizations directly. It had emailed three by the time it published.
It withheld every app address and title from the report.
How to Check Your Own Organization
Claroc's recommended checks are practical.
- Find out which AI app builders your staff uses. Grok, Lovable, Replit, v0, Bolt, Base44, and Google AI Studio can all be used without IT setup or a company account.
- Search passive DNS for the builders' domains, if you have access to that data. Open any results in a private browser window.
- Check every Grok app with internal data. It should be set to "Invited users" or "Only you," and a stranger should get an error page.
- Try signing in from an outside account. If an outside Google account gets in, the app needs an explicit allowlist.
- Review the scripts each app loads for names, email addresses, and passwords. Unpublish apps that have no owner.
In the researchers' tests, setting an app to "Only you" returned a 401 error to a stranger for both the page and its data. Unpublishing took an app offline within seconds.
This Is the Second Time
On October 18, the same researchers published Muse research showing that pages shared from Meta's Muse AI agent could be found through passive DNS.
Meta disabled sharing, then announced a fix on October 22 that gave shared items new addresses.
The pattern is consistent. AI app builders default to public, generate subdomains that leak through DNS, and the exposure is discoverable at scale.
FAQ
What did Claroc find?
37,571 Grok-built apps online, of which 35,578 were open to anyone without a sign-in. Twenty-four were built by staff at X and SpaceXSI.
What kind of data was exposed?
Staff apps contained a confidential onboarding plan, a sales territory map with 594 accounts, and an advertising proposal to a U.S. federal agency. Other open apps contained internal data from 30 identifiable organizations, including a hospital employee list, an ICU patient briefing, and a fire department shift board.
How were the apps found?
Through passive DNS. When someone opens a grok.me subdomain, DNS lookups are recorded by passive DNS services, and anyone with an account can search those records by pattern.
What is the default access setting?
"Anyone with the link" was already selected in the paid account the researchers used. ChatGPT's builder defaults to "Only me."
Did Grok's own sign-in protection work?
No. In the researchers' second test, the generated code added every signed-in user to the team before checking membership, so the check always passed. An uninvited Google account got full access.
Did SpaceXSI respond?
The company had not replied to the researchers' report by publication, but it answered a reporter's questions. It said apps stay private until shared and that employee accounts default to "Only you." Twenty-four staff apps were still open.
What did the researchers ask for?
Making "Only you" the default, and adding an option to restrict an app to a company's own email domain.
How do I check my own organization?
Find which AI builders your staff uses, search passive DNS for their domains, open results in a private window, verify access settings on any app with internal data, and try signing in from an outside account.
What happens when I set an app to "Only you"?
In the researchers' tests, it returned a 401 error to a stranger for both the page and its data. Unpublishing took the app offline within seconds.