Security researchers at Claroc revealed on November 11 that a Grok Build leak exposed thousands of web apps on the public internet [1]. The SpaceXSI Grok Build tool, built after Elon Musk merged SpaceX and xAI, lets users turn chat prompts into live web tools within minutes [3]. However, a flaw in how the service publishes software allowed passive DNS monitoring to discover private project links, while default sharing settings left almost all generated tools open without passwords [1]. The resulting exposure uncovered business records, hospital files, and 24 internal applications belonging to SpaceXSI itself [2].
How the Grok Build Leak Started
Claroc uncovered the flaw while investigating how the AI coding system assigns web links to newly generated projects. When a user builds an app, the platform creates a web link formatted as four random dice words under the grok.me domain, an arrangement designed with up to 38 bits of entropy to prevent guessing. However, as reported by TechRadar reporter Sead Fadilpašić, SpaceXSI placed those randomized words as a subdomain to the left of the domain name rather than as a path to the right, which Google Gemini does to keep links unlisted [1, 2]. That choice broke privacy [2]. Every project URL became easy to find through standard network lookups [3].
Because internet routing sends subdomains through lookup requests, passive DNS sensors logged every new grok.me address in public registries. The Claroc team gathered these records and found 43,601 distinct links created through the coding tool [1]. The researchers called these open pages “GhostShares,” because anyone scanning DNS registries could find and launch them without ever getting a link from the creator. Unlisted software became public [2].
Claroc then checked whether these discovered links enforced any visitor authentication, finding that 35,578 projects, representing 94.7% of all cataloged grok.me sites, opened completely without requiring a password or an account login. Security was absent. Only one solitary project in the entire collection, an archive of ad measurement studies run by a streaming service, asked visiting users to provide a password before viewing stored records. Everything else sat wide open [1].
Flaws in the Subdomain Design
Grok Build lets people build web apps simply by describing what they want in a chat window. SpaceXSI writes the code, hosts the files, and offers three privacy levels: Anyone with the link, “Invited Users”, and “Only You”. This setup mimics Google Docs. But SpaceXSI chose “Anyone with the link” as the default setting, so users in a hurry published without restricting access. Most people accepted the default [1].
To test whether outsiders could edit stored data, Claroc built two test apps on the platform. The first test proved that unauthenticated visitors could alter entries in a shared contact database, showing that outsiders could overwrite files without leaving an audit trail. The second test revealed that signing in with any personal Google account automatically gave full admin rights, even when that visitor had no ties to the project team. These tests proved that the flaw went far beyond passive viewing. Strangers could inject bad data [1].
The root cause lay in how domain name servers handle queries for new hostnames. Because each app lived on its own subdomain under grok.me, every lookup was recorded by passive DNS networks [1, 2]. If SpaceXSI had placed apps under URL paths on a shared root domain, external DNS logs would never have exposed the links. Subdomains exposed the files [2].
Medical Records and Corporate Files
Most exposed apps were harmless tools like games or basic calculators. However, roughly one in ten projects held internal business files, customer details, or employee rosters from active firms, with Claroc calculating that 2% of the open tools stored material serious enough to warrant a critical risk rating [1, 2]. Big companies suffered exposure. The exposed records affected six firms in the Fortune 500 list, while 26 impacted employers had workforces exceeding 1,000 workers across logistics, healthcare, and retail sectors [2].
The leaked data varied widely between groups that adopted Grok Build for quick workplace solutions [1, 3]. Claroc found a hospital staff list containing more than 3,500 names and contact numbers. Another open app held an intensive care briefing with patient records and clinical exam notes, exposing private health details without any login prompt [1, 3]. A recruitment database from a trucking firm showed personal files for more than 1,000 job candidates. In another case, an employee timesheet had downloadable code holding 18 accounts with plaintext passwords, giving anyone admin access [1].
As reported by Steve Dent for Engadget, data also leaked from a municipal fire department and a luxury cruise line that left galley checklists online [2, 3]. Workers built these tools by pasting shift spreadsheets into Grok chat prompts to save time [1, 3]. They never suspected that Grok Build published those private records to open web links [1].
What the Grok Build Leak Exposed at SpaceXSI
The most striking discovery in the Grok Build leak was that SpaceXSI exposed its own internal documents through the platform. Claroc identified 24 open applications created by SpaceXSI staff that held confidential company files [1, 2]. “We were shocked,” the Claroc research team wrote in a blog post. “The company that makes Grok had its own internal material open to anyone holding the link,” the researchers said [2]. Among the exposed files was a private orientation packet prepared for a newly hired representative joining the enterprise sales group at SpaceXSI [3].
As reported by Ars Technica, the SpaceXSI apps contained sales plans and client records that were never meant for outside eyes, revealing how employees at the Elon Musk-owned company shared sensitive files across open grok.me web links without password checks [2, 3]. The exposed sales files were extensive. The Claroc team found a sales territory map detailing 594 strategic accounts and pipeline figures, along with internal system web addresses used by sales staff during daily client outreach operations [2, 3]. Other open tools held eight ad proposals for X, four active ad partner dashboards alongside two separate versions of a brand safety guide for X, and a confidential media proposal prepared for a US federal agency [1, 2]. Seven international campaign decks also leaked [1].
The leak showed how internal teams bypass security reviews when using conversational AI to create custom tools. Employees used the chat interface to build quick administrative trackers that ended up on public grok.me subdomains [1]. The incident follows earlier operational adjustments across xAI and SpaceX, where Elon Musk capped Grok bot usage limits to control ballooning server and inference expenses. Internal convenience produced an embarrassing security exposure [2].
Timeline and Take Down Response
Claroc notified SpaceXSI about the widespread security exposure on November 8 [1]. But the company didn’t respond to the private notice. The Claroc team only gained traction after publicly tagging Elon Musk and another company leader on X. Within two days of that post, SpaceXSI took down all 24 of its exposed apps, cutting off public access to its internal sales maps and employee guides [1, 2].
Although SpaceXSI removed its own exposed applications, it did not take down the thousands of projects left open by other users. “It has not done the same for the apps other Grok users and organizations left open,” Claroc researchers wrote in their published report. Because SpaceXSI left outside tools online, Claroc began working through the database to contact affected firms, hospitals, and fire departments directly [1, 2]. Several companies acted quickly to shut down their apps after receiving warnings [2].
Claroc researchers warned that AI coding tools create fresh hazards when platforms trade security for speed. When automated systems set discoverable subdomains and default to open sharing, anyone watching DNS traffic can read private data without typing a password [1]. The Grok Build leak shows that quick app building can’t replace basic privacy protections when handling business files.
- ONLINE NEWS Fadilpašić, S. (2026, November 11). Elon Musk’s Grok Build coding agent is leaking a serious amount of confidential data. TechRadar. [Article Link]
- ONLINE NEWS Dent, S. (2026, November 11). A Grok vibe-coding agent leaked sensitive SpaceXSI data. Engadget. [Article Link]
- ONLINE NEWS Ars Technica. (2026, November 11). Grok Build coding agent leaks oodles of confidential info, some belonging to SpaceXSI. Ars Technica. [Article Link]
