AIO Apex
en

Grok Build Set 35,000 Apps Public by Default, Leaking SpaceX Data and Patient Records

The Hacker News
Share:
Grok Build Set 35,000 Apps Public by Default, Leaking SpaceX Data and Patient Records

When xAI opened its Grok Build app creator to all users in August 2026, it shipped with a quiet default that turned out to matter enormously: every published app was set to «Anyone with the link» — and those links were trivially discoverable.

Security firm Claroc catalogued 43,601 live grok.me hostnames using passive DNS, then tested 37,571 of them. Of those, 35,578 — 94.7% — opened for anyone without a password, as first reported by The Hacker News on November 11, 2026.

What Was Inside

The scale of the exposure wasn't just an access-control abstraction. Among the open apps, Claroc found 24 built by employees of SpaceXSI and X. Those included a sales territory map covering 594 enterprise accounts, eight advertising proposals (one addressed to a US federal agency), and a confidential onboarding plan for a new enterprise sales hire.

Beyond SpaceX and X, the firm identified 30 other organisations spanning six Fortune 500 companies. A hospital's staff roster of more than 3,500 people was open, alongside an ICU briefing that included one patient's lab results and imaging. A trucking company's recruiting list of over 1,000 candidates was accessible, as was a contractor app whose client-side code contained plaintext passwords for 18 accounts — including administrator accounts.

Claroc estimated that roughly 650 to 800 of the exposed apps contained data rated high or critical severity.

Not a Hack — a Design Choice

No authentication systems were bypassed. The exposure came from two compounding design decisions.

First, Grok Build's sharing step pre-selected «Anyone with the link» rather than a private default. Second, the grok.me subdomains assigned to each app were logged by passive DNS services, making them searchable by anyone with access to those feeds — regardless of whether the creator had ever shared the link.

Claroc also found a third layer: AI-generated authentication code that appeared correct but was not. In one test app, the sign-in logic added every Google-authenticated visitor to the team before checking team membership, meaning any Google account could read, edit, and delete records. In another, a database table had no authentication at all, letting anonymous visitors append rows that other users could then view.

The Response

Claroc notified xAI on November 8, 2026, initially with no reply. After the firm posted on X tagging Elon Musk and Grok product manager Matt Palmer — and after Ars Technica began asking questions — SpaceXSI removed its 24 internal apps within hours.

Third-party apps belonging to hospitals, contractors, and other businesses remained accessible. At the time of publication, xAI had not announced a plan to notify those app owners, nor had it committed to changing the sharing default.

A Wider Pattern

The Grok Build findings fit a broader pattern emerging from AI-assisted development: code generators produce auth logic that compiles and runs, but does not enforce the access rules the developer intended. The Cloud Security Alliance, in a research note published the same day, warned that «generating authentication with AI does not guarantee that the authorization logic is correct.»

As AI app builders move from early adopters to enterprise teams, the defaults they ship with — and the security review gap between «the AI wrote it» and «a human audited it» — may leave far more than 35,000 apps in the same position.

Originally reported by The Hacker News. Read the original article for additional details.

View original source
Share: