Thousands of Grok Apps Exposed Corporate and Personal Data Without Sign-In
November 11, 2026
A Claroc Research review published November 10 found that 35,578 of 37,571 Grok Build apps it tested could be opened without signing in, exposing confidential corporate, employee and personal information; 24 were built by SpaceXSI and X staff.
Among the exposed material were business plans, advertising proposals and a database of 594 strategic customer accounts, with information also linked to at least 30 other identified organizations.
Claroc traced the exposure to shareable app links that seemed difficult to guess but were discoverable through public passive DNS records, which can capture hostnames when a website is opened.
Grok’s links use randomly generated four-word phrases, but place them before the domain in the hostname, where passive DNS services can record and searchable databases can expose them; by contrast, secret link components in Claude, ChatGPT and Gemini URLs sit in the path, not the DNS-visible hostname.
The dispute is whether users simply chose permissive sharing settings or whether Grok’s publishing interface and hosting design made apps marked for link-based sharing discoverable in practice.
SpaceXSI said apps stay private until creators share them and employee accounts default to “Only you”; Claroc said the publishing panel preselects “Anyone with the link,” creating risk, and urged private-by-default settings and harder-to-discover URLs.
Claroc’s tests found that “Only you” blocked outside access, and it recommended organization-based access controls, including restrictions by email domain.
Claroc said it demonstrated the issues only in its own test apps and did not alter third-party data; researchers could not establish whether anyone malicious had accessed exposed information.
In tests on researcher-created apps, Claroc also found code-level access-control failures that could let outsiders add records or let an uninvited account view, change or delete fictional customer data; it said these tests did not involve other users’ apps.
Claroc advises organizations to inventory AI app builders, search for exposed app addresses, verify sign-in and sharing boundaries, review code for embedded secrets, test access from an outside account and take down apps without clear owners; its tests found “Only you” blocked strangers and unpublishing took apps offline within seconds.
The incident highlights how AI app builders can turn prototypes into internet-facing services quickly, while users may not understand permissions, network visibility or the security implications of publishing.
The concern extends beyond Grok Build: tools including Lovable, Replit, v0, Bolt, Base44 and Google AI Studio can enable employees to publish internal tools without IT review, though exposure depends on each app’s settings and implementation.
Summary based on 26 sources
Get a daily email with more Tech stories
Sources

Ars Technica • Nov 10, 2026
Grok Build coding agent leaks oodles of confidential info, some belonging to SpaceXSI
TechRadar • Nov 11, 2026
Elon Musk's Grok Build coding agent is leaking a serious amount of confidential data
Engadget • Nov 11, 2026
A Grok Vibe-Coding Agent Leaked Sensitive SpaceXSI Data
The Hacker News • Nov 11, 2026
Over 35,000 Grok-Built Apps Exposed Online, Some Leaking Confidential Data and Passwords