Security researchers found that default settings in Elon Musk’s coding agent exposed thousands of private web apps to the public.
A misconfiguration in Grok Build, an artificial intelligence tool developed by Elon Musk’s company SpaceXSI, exposed tens of thousands of web applications containing private corporate records, healthcare files and administrative passwords to the internet, according to cybersecurity researchers at Claroc.
In a report detailed by TechRadar, researchers explained that Grok Build functions as an AI coding agent allowing people to generate web applications by describing what they want in a chat window. SpaceXSI, created through a merger between SpaceX and xAI, lets users publish their software directly from that interface without drafting any code. Each project receives an automated web address ending in grok.me, preceded by four randomly generated words designed to keep outsiders from guessing the location.
How web addresses were exposed
Engadget reported that the platform’s naming scheme inadvertently made those addresses discoverable. Grok Build placed the randomly generated four-word phrase to the left of the grok.me domain name rather than to the right, unlike systems such as Google Gemini. That structure allowed passive domain name system services, which monitor and index traffic on the web, to log and catalogue the addresses into searchable databases.
Claroc discovered 43,601 distinct web addresses under the grok.me domain by searching passive DNS records. The investigation revealed that 35,578 of those applications, or 94.7 percent, were fully accessible without any authentication requirement. The researchers coined the term GhostShares for these unprotected pages because anyone with access to internet traffic logs could locate and open them without ever receiving an invitation or link from the creator.
The lack of access protection stemmed from how Grok Build configured sharing choices. When people deploy an application, the interface presents three options: Anyone with the link, Invited Users, and Only You. SpaceXSI set Anyone with the link as the standard default setting, meaning most creators published their work publicly without adjusting their permissions.
Medical records and company passwords
While many of the exposed projects were harmless tools like calculators, landing pages and casual games, Claroc reported that roughly one in ten applications contained internal business details or sensitive private records. About two percent of all reviewed projects contained data severe enough to qualify for a high or critical risk rating.
The unsecured applications included a hospital employee directory with more than 3,500 workers, an intensive care briefing document listing confidential patient information, and a recruitment database for a trucking firm with records for over 1,000 people. Claroc also discovered exposed records linked to a local fire department.
In another instance, an employee timesheet application contained downloadable source code that exposed 18 different user accounts with passwords stored in plaintext, including login credentials for system administrators. The leak also touched major corporations, including six businesses listed on the Fortune 500 and 26 firms with workforces exceeding 1,000 employees.
SpaceXSI exposed its own internal files
The exposure extended to SpaceXSI itself. Researchers uncovered 24 public applications created by the company that lacked basic password requirements. Only a single SpaceXSI application found by the team, which held an archive of advertising measurement studies for a streaming service, required a password.
“The company that makes Grok had its own internal material open to anyone holding the link,” Claroc wrote in a blog post describing the breach.
The exposed SpaceXSI applications contained staff onboarding schedules, rosters of employee names, four live advertiser account boards, eight advertising proposals for social network X, two copies of an X brand safety playbook, and seven campaign reports for commercial clients worldwide. The files also contained a sales territory mapping document listing 594 strategic accounts alongside their business pipelines, as well as confidential advertising pitch decks detailing specific budget tiers, including one prepared for a United States federal agency.
During testing, Claroc confirmed that visitors could actively modify data in some applications. The researchers built two test apps to evaluate the system. One let unauthenticated visitors alter a shared contact list, while the second granted access to any user logging in with an outside Google account regardless of team membership.
Notification and response
Claroc reported that it notified SpaceXSI about the security vulnerability on November 8. SpaceXSI did not respond immediately until Claroc reached out on X by tagging Musk and another company representative.
Within two days of the notice, SpaceXSI took down all 24 of its exposed applications. Other affected businesses alerted by Claroc also took quick steps to restrict their unsecured pages. However, Claroc stated that SpaceXSI did not implement a universal fix to close the remaining exposed applications built by other customers and organizations, leaving researchers to work through the database to notify affected third parties directly.
