Clarity for AI buildersSponsored by AI Routing
adoption

Claroc finds 35,578 Grok-built apps accessible without login, including sensitive records

Claroc finds 35,578 Grok-built apps accessible without login, including sensitive records
Claroc found 35,578 Grok-built apps accessible without login, The Decoder reports, including apps exposing sensitive data. According to The Decoder’s account of the security firm’s findings, 35,578 of 37,571 still-online apps checked by Claroc were accessible to anyone—94.7 percent of that sample. Exposed material included patient information, plaintext passwords and internal business documents. The findings concern published apps in the examined sample, not every app created with Grok. Grok’s Build mode lets users describe an app in a chat and have the assistant write its code. The Decoder identifies the provider as SpaceXSI, formerly xAI. Apps remain private until their creators publish them, giving each one an address on grok.me. On the paid account Claroc tested, however, the publishing dialog had “Anyone with the link” preselected. Claroc also found that app addresses were discoverable through commercial passive DNS databases. Each published app receives a separate subdomain, which the report says appears in those databases after its first visit. This creates a distinction between an address that a creator has not deliberately circulated and an app that actually restricts access: a discoverable link offers no access barrier when login is not required. Sensitive information among mostly ordinary apps Most accessible apps were games, landing pages or calculators, according to Claroc. The firm said about one in ten contained sensitive material, internal business data or personal information, and classified roughly two percent as especially serious. It separately estimated that 650 to 800 apps contained private data; those categories should not be treated as interchangeable counts. Reported examples included an ICU briefing containing a patient’s laboratory values, a list of more than 3,500 hospital-group employees and a fire department shift schedule. A timesheet app’s source code contained 18 accounts with plaintext passwords. Beyond SpaceXSI and X, Claroc identified 30 affected organizations, including six Fortune 500 companies, but withheld their names. The firm attributed 24 accessible apps to SpaceXSI and X employees. These included a confidential onboarding plan with internal system addresses and 19 chat channels, a sales territory map covering 594 strategic accounts and pipeline data, and eight X advertising proposals. One proposal, aimed at a U.S. federal agency, was marked confidential. Login did not always enforce team membership Claroc’s own tests identified a separate authorization problem. In a sales customer-list app, Grok added Google and X sign-in, but the generated code enrolled authenticated visitors into the team before checking membership. An outside Google account could consequently read, edit and delete records. Claroc said it performed write and delete tests only on its own apps. That test illustrates why requiring sign-in and restricting access are different controls. Authentication established that a visitor had an account; the flawed membership sequence failed to establish whether that visitor belonged to the intended team. A separate contact-list test also produced a database that anonymous visitors could add entries to. Response and limits SpaceXSI said apps are private until shared, publishing requires multiple steps including a legal notice, and employee accounts default to “Only you,” according to The Decoder. Claroc countered that additional steps do not resolve a public-access option being preselected. Its test of “Only you” successfully blocked outside access, indicating that the availability of a restrictive setting and its use are separate issues. The report says SpaceXSI took the 24 employee apps offline, while other users’ exposed apps remained available. Claroc proposed private-by-default sharing, clearer audience warnings and warnings for databases without login. It also recommended shared-host paths rather than individual subdomains to reduce DNS-based discovery; that change would address discoverability, not replace access controls. The supplied evidence is The Decoder’s account, rather than Claroc’s underlying report or an independent retest. The publication also notes that Claroc sells security products for AI-built apps, a commercial interest relevant to assessing its findings.