On this page
Grok Build leaks sensitive user data, according to security firm Claroc, which said on 10 November 2026 that it could open 35,578 apps built with the coding agent without signing in. Some of those apps held medical, staff and business information. The company behind Grok Build, which Claroc calls SpaceXSI (formerly xAI), took down 24 apps built by its own staff after the report but, according to The Next Web, left other users' apps open.
This is a different problem from the July 2026 episode in which Grok Build uploaded users' code repositories. Both are covered below, because the two are easily confused.
What did Claroc report on 10 November 2026?
Claroc published its findings in a research post dated 10 November 2026. The firm said it tested 43,551 Grok Build app addresses. Of the 37,571 still online, 35,578 (94.7%) opened for anyone, with no sign-in. The other 1,993 returned an HTTP 401 response, which means they asked for credentials.
Most of the open apps were harmless: Claroc says they were mainly games, landing pages and calculators. But it estimates that about one in ten held sensitive material, which it puts at roughly 650 to 800 apps.
| Claroc figure | Value |
|---|---|
| Addresses tested | 43,551 |
| Apps still online | 37,571 |
| Apps open without sign-in | 35,578 (94.7%) |
| Apps built by SpaceXSI and X staff | 24 |
| Other identifiable organisations affected | 30 (26 with over 1,000 employees, six in the Fortune 500) |
Claroc did not name the other organisations. These are the firm's own numbers, and we have not been able to check them independently.
What data was exposed?
According to Claroc, the open apps included:
- A hospital staff roster, an intensive care patient briefing and a fire department shift board.
- A staff timesheet that stored employee passwords in plain text.
- From SpaceXSI and X staff: a confidential onboarding plan, a sales territory map covering 594 accounts, and advertising proposals, one marked confidential and addressed to a US federal agency.
Coverage by Dataconomy and Startup Fortune repeats the same categories of material. Both rely on Claroc and on earlier reporting by Ars Technica rather than on their own testing.
Why were the apps findable?
Claroc says the publish panel in its paid test account came with "Anyone with the link" preselected, so public access was the default. It also says each published app gets its own hostname, and that hostnames can show up in passive DNS records, which are searchable. In practice that means an app could be found without anyone having been sent its link.
In plain terms, a link that is not advertised is not the same as access control. Claroc's recommendation is that "Only you" should be the default, that shared apps should use hard-to-guess paths, and that builders should be warned when an app has a database but no sign-in. Claroc also says it built two test apps of its own, and that one of them had a flaw that let outside accounts read, edit and delete records. We are not describing how that works.
What has SpaceXSI said?
According to The Next Web, which quotes Ars Technica, the company's representative said: "Apps are private by default and nothing goes public unless the creator chooses to share it." The company also said publishing involves several steps, including a legal notice, and that employee accounts default to "Only you." The same report says the company pointed out that other tools, including ChatGPT, Lovable and Bolt, also offer shareable links.
Claroc disputes the "private by default" claim. It says the access step it saw was preselected to "Anyone with the link", and that the people who published the open apps went through the extra steps anyway.
Claroc's own disclosure timeline says it reported the issue on 8 November, reported the 24 staff apps by 9 November, and that on 10 November SpaceXSI answered a reporter's questions and removed those 24 apps. We could not find a public SpaceXSI or xAI statement on the other users' apps, nor any confirmation that the default has changed. We have not opened the Ars Technica article ourselves, so its content is taken from The Next Web's account.
How is this different from the July Grok Build upload story?
In July 2026, researchers at Cereblab found that Grok Build was packaging users' whole repositories and uploading them to cloud storage. The Register reported on 16 July 2026 that SpaceX said retention had been enabled by default for non-enterprise users during the early beta, that the default was switched off from 12 July, and that previously retained coding data was being deleted. The company then released the tool's code on GitHub.
That was about what the tool sent to the company's servers. The November findings are about what apps users publish with it. Different mechanism, same product.
Analysis: what does this mean for teams using AI app builders?
This section is our own analysis, not reporting.
The common thread is defaults. Whichever side is right about "private by default", the dispute shows how much depends on a single preselected option at the moment of publishing. Claroc itself says rival builders have similar exposure, so the lesson is not limited to one vendor. Companies that let staff build internal tools with AI agents may want to find out which builders are in use and what each one's sharing options are. We covered a related kind of exposure in the Notion AI chat transcripts incident, and the wider shift towards autonomous tools is explained in our guide to agentic AI.
Sources
- Claroc Research: Grok-built apps exposed internal and personal data (10 November 2026)
- The Next Web: Grok Build left 35,000+ apps open, Claroc says (12 November 2026)
- Dataconomy: Grok Build leak exposes sensitive data from thousands of users (11 November 2026)
- Startup Fortune: Grok Build links leaked SpaceXSI sales data and thousands of customer apps
- The Register: SpaceX open-sources Grok Build after data retention furore (16 July 2026)
FAQ
Frequently asked questions
Claroc says apps built with Grok Build exposed sensitive data to anyone who found them, including a hospital staff roster and employee records. The company disputes that apps are public by default. The figures are Claroc's own and are reported by The Next Web, Dataconomy and Startup Fortune.
Claroc says 35,578 of 37,571 live apps it tested opened without sign-in. It estimates about 650 to 800 of them held private data.
SpaceXSI removed 24 apps built by its own staff on 10 November 2026. According to The Next Web, other users' apps were left open at the time of reporting. We found no public confirmation of a wider fix.
No. The July issue was about the tool uploading users' repositories to cloud storage. The November report is about apps published through share links.
Spotted an error? Tell us. Read how we report and correct stories in our editorial policy.
Written by
Mainstream Tech
Newsroom
News and analysis on technology, AI, business, politics and international affairs, from our newsroom in London.
More from Mainstream



