# Grok Build data leak exposes risks in AI-built apps > The Grok Build data leak exposed more than 35,000 apps through shareable links. Learn what was exposed and how teams should check AI-built apps. - URL: https://ferdousazad.com/blog/grok-build-data-leak - Author: Ferdous Azad (Founder @ DataXgen AI) - Published: 2026-11-11 - Updated: 2026-11-11 - Tags: ai, security, agents, ai-ethics ## 1. Introduction On November 11, 2026, Ars Technica reported that researchers had found more than 35,000 Grok Build apps accessible through shareable links, including apps containing confidential business information and personal data. For developers and founders, the warning is direct: an app that’s quick to build with an AI agent still needs deliberate access controls before it holds real company or customer data. ## What happened Security firm Claroc found that Grok Build, a coding agent from SpaceXSI, exposed data through links intended to let users share their apps. The URLs were randomized, but a design flaw in the platform made it possible to enumerate thousands of live projects. In most cases, someone who reached a link could open the app and view its associated data without company sign-in. The report, published by Ars Technica on November 11, says the exposed information included internal business plans, advertising proposals, medical exam results, employee names and ID numbers, and operational documents. Claroc also found 24 apps built by SpaceXSI itself. Examples included a confidential sales-team onboarding plan, a territory map covering 594 accounts, internal system addresses, and advertising proposals. Other affected app data included medical exam records for 3,784 workers at a Peruvian construction company, a hospital staff roster listing about 3,565 employees, and a Tesla service-center scorecard naming 22 technicians. Claroc also reported a roster connected to a US Navy logistics contractor and documents relating to an aerospace and defense group. According to Ars Technica, Claroc notified SpaceXSI on Sunday. When the company hadn’t responded, a Claroc official publicly raised the issue on X. SpaceXSI began removing links to its 24 apps a few hours after Ars sent questions; by publication, those links returned an “app not found” error. Ars reported that almost all other exposed data remained available at that time. ## Why it matters The central problem is the difference between a hard-to-guess URL and a real access-control boundary. Randomized links can make casual discovery less likely, but they don’t establish who is allowed to see an app. If a platform lets people enumerate links, that assumption breaks down. And even without enumeration, a link can be forwarded, copied into a ticket, or exposed in a browser history. The people at risk aren’t only the developer who created an app. The reported data could affect employees whose names, IDs, or health records appeared in dashboards, customers whose proposals included commercial details, and workers connected to industrial or defense operations. A company can expose sensitive information through a prototype built by one person, even when the people named in it never knew the app existed. This report also points to a common failure in AI-assisted development: generating an app can be fast, while deciding where its data lives and who can access it remains a separate engineering task. The facts reported so far point to a sharing and access-control failure. They don’t establish that the AI model itself caused the exposure. ## What it means for you - **Treat shared links as public unless access is verified.** Don’t put personal, regulated, or confidential data in an app just because its URL looks difficult to guess. Check what an unauthenticated person can see. - **Inventory AI-built apps and their data.** Find prototypes and internal tools created with coding agents, then identify what data they store or display and who can reach them. Informal tools are easy to miss in a normal service inventory. - **Review access and sharing controls before launch.** Confirm that sensitive apps require authentication and that permissions match the audience. Test from a private browser session or another account rather than relying on the creator’s logged-in view. - **Have an incident plan for exposed data.** If a link was public, assess what it revealed, revoke access where possible, and rotate any exposed credentials. Consider whether affected employees, customers, or partners need to be notified under your policies or obligations. ## The bigger picture AI coding agents can make it easier for teams to build small internal apps without a conventional software project. That can be useful, but it also means more people may create tools that handle company data without going through the usual review process. The security boundary still has to be designed and checked, whether the code was written by a developer or generated by an agent. This incident is a reminder that “shareable” and “private” are separate product decisions. Link randomization can help keep a URL from being guessed, but it shouldn’t substitute for authentication when the content is sensitive. For builders evaluating AI app platforms, the useful question isn’t only whether the tool can create an app; it’s whether the app’s data and permissions behave safely when the link leaves the creator’s control. ## Key takeaways - Ars Technica reported on November 11, 2026, that researchers found more than 35,000 Grok Build apps accessible through shareable links. - Claroc reported exposure of business information and personal data, including material from 24 SpaceXSI apps. - A randomized URL isn’t a reliable access-control mechanism, especially if links can be enumerated. - Audit AI-built tools, test access without an authenticated session, and keep sensitive data out of apps without verified protections. ## Sources - https://www.techmeme.com/261111/p3#a261111p3 - https://arstechnica.com/security/2026/11/grok-build-coding-agent-leaks-oodles-of-confidential-info-some-belonging-to-spacexsi/