Grok Build: 35,000 apps exposed, links found via passive DNS
Grok Build apps were shared through links in the subdomain: passive DNS revealed more than 35,000 of them open to anyone, with confidential business data.

More than 35,000 applications built with Grok Build, the “vibe coding” agent from SpaceXSI (the company formed by the merger of SpaceX and xAI), were reachable by anyone on the internet, with confidential business documents and personal data inside. The discovery comes from security firm Claroc, which published it on 10 November 2026; the story was picked up by Ars Technica, The Hacker News, DataBreaches.net and heise. There is no CVE: this is a design decision. The secret meant to protect the sharing links sat in a part of the address that DNS makes public.
How sharing worked
Grok Build generates a working app from a spoken or written description and lets you share it with a link. According to Ars Technica, the address takes the form <random-words>.grok.me: two or more randomly generated words (“dice words”, which users can also edit) before the domain. This is the capability URL model: whoever knows the address is authorised, so the address must be impossible to guess. With three random words, again according to Ars, entropy is roughly 38 bits.
The problem is where the secret sits. Services that use this model correctly, such as Google’s shared documents, put the random part in the URL path, after the domain. In Grok Build it sits in the hostname, to the left of grok.me. When someone opens the link, their resolver asks DNS for the address of that name, and passive DNS services record never-before-seen names in searchable databases. In practice, the list of shared apps was available to anyone. More generally, the hostname also travels in clear text in the SNI field of the TLS connection unless ECH is used, and ends up in proxy and firewall logs: it is no place for a secret.
In MITRE ATT&CK terms, the reconnaissance matches technique T1596.001 (Search Open Technical Databases: DNS/Passive DNS). After enumeration nothing else was needed: except in rare cases, opening the link showed the app with all its data.
What the researchers found
The figures reported by the sources differ slightly. According to Ars Technica, quoting Claroc CTO Asaf Aprozper, the passive DNS search returned 43,601 addresses under grok.me, of which more than 35,000 (nearly 95%) were public. The summary published by DataBreaches.net instead mentions 43,551 addresses examined and 35,578 open apps out of 37,571 active ones.
Among the exposed data, according to the list reported by Ars Technica: an occupational health dashboard with 3,784 worker medical exams from a Peruvian construction company, a roster of about 3,565 employees of a healthcare organisation with names and IDs, 108 people on a US Navy logistics support contract, the loading procedure for an explosives building belonging to a Norwegian-Finnish aerospace and defence group, confidential sales proposals and about 4,000 vehicle transport orders with prices and margins. The Hacker News states in its headline that some apps also exposed passwords.
SpaceXSI itself was among those exposed: Claroc found 24 of the company’s own apps, including a confidential onboarding plan for the enterprise sales team, a territory map of 594 accounts and the addresses of internal systems. According to Ars Technica, Claroc notified SpaceXSI on 8 November without getting a reply; the 24 apps were taken down a few hours after the outlet sent its questions, while the rest of the data was still reachable when the article was published.
Default setting: two versions
According to Aprozper’s tests, Grok Build sets links to “Anyone with the link”, and restricting them requires manually choosing “invited users” or “only you”. A spokesperson for Cursor, the AI coding company acquired by SpaceXSI, told Ars Technica instead that apps are private by default and only become public if their creator chooses to share them.
Claroc notes that other app-building platforms such as Lovable, Bolt and Replit also give each app its own hostname and can therefore be enumerated the same way, although it found far more sensitive business data in Grok. The public shares of Claude, ChatGPT and Gemini, by contrast, put the random part in the path, which DNS never sees.
Impact for European and Italian companies
The case is an example of shadow AI: tools built by an employee in minutes, with real data, outside the company’s access controls. If an app contained personal data of employees, patients or customers, its exposure is a breach for which the company is responsible as controller: it must be assessed under Articles 33 and 34 of the GDPR, with notification to the supervisory authority within 72 hours unless a risk to individuals is unlikely. For NIS2 entities the issue falls under the management of risk related to suppliers and the tools in use; in healthcare, clinical data make the exposure particularly serious. The sources do not mention any Italian companies being affected.
What to do now
- Find usage: check DNS, proxy and firewall logs for requests to
*.grok.meand to the domains of other app-building platforms, to see who uses them and for which projects. - Restrict or delete: for apps you need, set sharing to “invited users” or “only you”, as Ars Technica advises; delete the ones you no longer use.
- Assume the data has been read: changing the setting does not undo past access, because the addresses are already in passive DNS databases. Change any passwords and keys stored in the apps and assess whether you must notify.
- Clear rules: define which AI tools may be used, ban real personal data in prototypes and require sign-in with the company account (SSO) for anything that gets shared.
- If you build services with sharing links: put the token in the path, not in the subdomain, and pair it with access control when the data is confidential.
Last updated:


