A default setting that quietly became the problem
Claroc's scan of 43,551 grok.me web addresses found 37,571 active Grok Build apps, and of those, 35,578, nearly every single one, were reachable by anyone with the link, requiring no login or authentication whatsoever. That ratio is the headline finding: this was not a small number of careless users misconfiguring their individual apps, it was the overwhelming default behavior of the platform itself, which means the exposure pattern reflects a product design choice rather than a scattered set of user errors.
SpaceXSI's own response to the findings confirms this directly: the company said Grok apps stay private until a creator actively chooses to share them, and that its publish option offers only two settings, restricted to the creator alone, or open to anyone holding the link. There is no intermediate option, no way to share an app with a specific team or organization without making it reachable by literally anyone on the internet who obtains the URL, which is a meaningfully narrower set of choices than most enterprise-grade sharing tools offer by default.
What was actually sitting behind those open links
Claroc found roughly one in ten of the publicly accessible apps contained something genuinely sensitive: internal corporate files, advertiser dashboards showing real campaign spending figures, a confidential proposal addressed to a US federal agency, a searchable roster of more than 3,500 hospital employees, and an ICU patient briefing containing actual lab results. A cruise ship crew list with employee identification numbers and a fire department's internal shift board were also among the exposed material, spanning healthcare, government, maritime, and public safety contexts that have nothing in common except that an employee somewhere used Grok Build to create a quick internal tool.
One specific app, built to manage contractor timesheets, stored plaintext passwords for 18 employees, including several with administrator-level access, which is a distinct and independently serious failure layered on top of the broader public-accessibility problem. A publicly reachable app is one level of risk; a publicly reachable app that also stores credentials in plaintext compounds that risk into something an attacker could use to pivot directly into other systems those 18 employees have access to elsewhere.
Why SpaceXSI and X employees appearing in their own leak matters
At least 24 of the exposed apps were built by employees of SpaceXSI and X, meaning the company behind Grok Build had its own internal material exposed by the exact same platform weakness it had shipped to every other customer. That is a notable detail beyond simple irony: it suggests the company's own employees were using the tool the same casual way external customers were, without any internal policy or technical guardrail steering them toward more restrictive sharing settings for sensitive internal material.
Claroc's researchers described being genuinely surprised that a company building and operating Grok itself had internal material openly reachable by anyone with the link, which is a stronger statement than a researcher typically makes about a vendor's own security posture. If the company building the tool was not applying additional internal caution to its own use of it, that is a meaningful signal about how little friction existed anywhere in the product's default sharing flow.
A fix that covers the company's own exposure and stops there
After Claroc's reports on November 8 and 9, SpaceXSI removed all 24 of the apps its own employees had built, which addresses the company's direct exposure but leaves the structural problem that created it entirely unaddressed for every other customer. Researchers specifically noted that the company had not taken equivalent action on the exposed apps belonging to the roughly 30 other affected organizations, several of them Fortune 500 companies, meaning those apps may still be sitting open to anyone with the link at the time of Claroc's report.
Fixing your own exposure while leaving an identical structural vulnerability live for every other customer is a meaningfully incomplete response to a platform-level finding like this one. A genuine fix would change the default sharing behavior for all users going forward, or at minimum notify every organization whose apps Claroc flagged as containing sensitive data, neither of which SpaceXSI appears to have committed to doing based on what researchers have described publicly.
Why this is a governance problem, not just a platform bug
Claroc's recommendation to organizations was specific and practical: audit which AI app-building platforms employees are actually using across the organization, and add safeguards against the kind of casual data leaks this incident represents. That recommendation matters because the underlying behavior, an employee spinning up a quick internal tool without routing it through any formal IT review, is exactly the kind of shadow-IT pattern that AI app builders make dramatically easier and faster than it used to be.
A tool that lets a non-technical employee build and publish a working internal application in minutes is genuinely valuable, but that same speed removes the natural friction that used to force a conversation with IT before something touching patient data, employee rosters, or federal proposals went live anywhere. The 35,578 publicly exposed apps Claroc found did not each represent a deliberate decision to expose sensitive data; they represent the absence of any decision at all about data sensitivity before publishing.
What this means for any enterprise whose employees use AI app builders
Any enterprise that has not explicitly addressed AI app-building tools in its data governance policy should treat this incident as direct evidence that employees are already using these tools to handle sensitive data, often without realizing the sharing defaults expose that data publicly by default rather than privately. The hospital roster, ICU briefing, and federal proposal Claroc found were not built by engineers deliberately testing security boundaries, they were built by employees solving an immediate, practical problem with the fastest tool available to them.
The practical response is not to ban these tools outright, which tends to push the same behavior further underground and out of IT's visibility entirely, but to establish clear guidance on what kinds of data are acceptable to put into a quick-build AI app platform at all, combined with periodic audits similar to the one Claroc just ran voluntarily and for free. Waiting for an external security researcher to find your organization's exposed hospital roster is a meaningfully worse way to discover this problem than finding it through your own internal review first.



