Grok Build’s Link Design Flaw Exposed 35,000+ Apps and SpaceXSI’s Own Secrets

A design decision buried in how Grok Build generates shareable links has turned tens of thousands of AI-generated applications into open doors. Security firm Claroc found that 35,578 of 37,571 live Grok Build apps could be opened without signing in, exposing sensitive material from hospitals, fire departments, Fortune 500 firms and SpaceXSI itself — the company behind Grok.

• Scale of the exposure: Claroc probed 43,551 grok.me addresses; 37,571 were still online and 94.7% of those loaded without authentication.
• Who was hit: At least 30 organizations outside SpaceXSI and X, including six Fortune 500 companies and 26 employers with more than 1,000 staff.
• Root cause: Random-word subdomains placed before the domain name, combined with a preselected ‘Anyone with the link’ sharing option, made private apps discoverable through passive DNS records.

What Actually Leaked

The exposed material ranged from the mundane to the alarming. Most of the open apps were games, calculators and landing pages, but Claroc estimated that roughly one in ten contained sensitive personal or business information, and about 2% held material rated high or critical.

Among the documented cases: an intensive-care briefing for a patient that included laboratory results, imaging records and ventilator usage logs; a contractor timesheet with the plaintext passwords of 18 accounts written directly into the script; and a Tesla service-center scorecard listing the names, efficiency ratings and bonus details of 22 technicians.

SpaceXSI’s own footprint was equally revealing. Claroc identified 24 internal apps built by SpaceXSI and X employees, including a confidential onboarding plan, a sales territory map covering 594 strategic accounts, eight advertising proposals for X, four live advertiser account boards and two copies of an internal brand-safety playbook. One advertising proposal was addressed to a U.S. federal agency.

How a Random Link Became a Public Index

Grok Build assigns every published app its own grok.me hostname, built from a string of random words intended to be hard to guess. The flaw was placement: the random portion sat to the left of the domain, where passive DNS services could capture it. Every visit to an app generated a DNS lookup, that lookup was logged, and the hostname became searchable in public databases. Claroc dubbed the resulting pages ‘GhostShares’.

No password was ever cracked. The ‘secret’ links simply left a public trail. In Claroc’s paid test account, the publishing panel already had ‘Anyone with the link’ selected — a default that effectively became policy for every builder who never changed it.

The researchers summarized their initial reaction in two words: ‘we were stunned.’

The comparison with competitors is instructive. Claroc ran a control test on ChatGPT-built apps: their addresses can also surface in passive DNS, but the sharing panel defaults to ‘Only me’, so a discovered URL leads nowhere. A Cursor representative likewise described its apps as private by default.

A Slow Response and a Governance Gap

Claroc notified SpaceXSI and the affected organizations. Some reacted quickly to contain the exposure. SpaceXSI did not respond until the researchers tagged Elon Musk and another company official on X. The company then removed its 24 internal apps but, according to the report, took no comparable action for apps built by other users and organizations, leaving hundreds of exposed apps online at publication.

SpaceXSI told a reporter that employee accounts default to ‘Only you’ and that apps remain private until creators choose to share them — a claim that conflicts with what Claroc observed in its own test account. The report documents public accessibility but does not establish that outside attackers stole or misused the data.

Claroc’s follow-up tests suggest the problem runs deeper than sharing settings. In one fabricated app, visitors could add records without signing in; in another, an uninvited Google account could read, edit and delete customer records — a broken authorization flow in which the app added the visitor to the team before checking permissions.

The broader lesson concerns the economics of ‘vibe coding’. When anyone can ship a data-connected app in minutes, security decisions migrate from specialist teams to whoever clicks create, and a single preselected toggle scales across tens of thousands of deployments. Fixing 24 apps addresses one company’s exposure; fixing the default addresses everyone else’s. Until AI app builders treat public sharing as an explicit, deliberate choice and enforce authorization server-side, the fastest path to software will remain the fastest path to unintended disclosure.