Claroc finds 94.7% of Grok-built apps were open to anyone with the link
Security firm Claroc checked 37,571 still-live apps built with xAI's Grok "Build" feature and found 35,578 of them, or 94.7 percent, could be opened by anyone with the URL, no account required, per Claroc's published findings.
What
Grok Build lets a user describe an app in chat and get working code back, with each published app getting its own address on grok.me. Claroc traced the exposure to two compounding defaults: "Anyone with the link" was pre-selected on the publish dialog, and each app's subdomain became searchable in commercial passive-DNS databases after its first visit, according to the report cited by The Decoder. Claroc estimates 650 to 800 of the open apps held private data, with confirmed cases including an ICU briefing listing a patient's lab values, a roster of more than 3,500 employees at an unnamed hospital group, a fire department's shift schedule, and a timesheet app whose source code contained 18 accounts with plaintext passwords. Separately, Claroc found 24 open apps built by xAI and X staff, among them a sales-territory map naming 594 strategic accounts and eight ad proposals, one of which targeted a US federal agency and was marked confidential.
In two of its own test apps, Claroc found a second flaw: Grok's generated code could add a database with no login check at all, and in one case it granted every newly signed-in Google or X account team membership before verifying that membership, letting an outside account read, edit, and delete records.
Login defaults now decide what a vibe-coded app leaks
For anyone building apps through a chat-based tool rather than writing auth logic by hand, the publish-time default is the entire security boundary, and Claroc found it set to the least private option. The firm noted that ChatGPT's equivalent sharing flow defaults to "Only me," a direct contrast that puts the choice of default, not just the existence of a privacy setting, at the center of this incident. SpaceXSI (the company formerly branded xAI) told Claroc that publishing already requires several steps, including a legal notice, and that staff accounts default to the stricter setting. Claroc countered that extra clicks do not matter if the final step still defaults to public, pointing to the 24 exposed staff apps as proof the stricter employee default did not actually hold.
This pattern is not unique to Grok, either. Claroc says it used the same passive-DNS technique in October to locate roughly 6,800 shared conversations from Meta's Muse agent, per the same Decoder report.
What to watch next
SpaceXSI has taken the 24 staff apps offline but, per Claroc, has not removed the exposed apps belonging to outside users and organizations, and has not committed to notifying them, which is why Claroc says it is contacting affected owners directly. Watch for whether xAI changes the publish-time default itself, and whether any of the 30 other affected organizations Claroc identified, six of them Fortune 500 companies whose names Claroc is withholding, confirm a breach.
Sources
- Claroc blog: GhostShares - Grok Build apps exposure findings: primary research
- The Decoder: Grok-built apps exposed patient data, plaintext passwords, and SpaceXSI's own sales pipeline: secondary confirmation
