Grok Build left 35,578 apps open — 94.7% of live projects require no login
Grok Build builds an app from a chat description and publishes it to grok.me, and by default it offers access to anyone with the link. Security firm Claroc checked more than 43,500 addresses: 37,571 apps were still running, and 35,578 of them (94.7%) opened without a login or password. Roughly one in ten exposed work or personal data. That is why anything built with an AI builder should be set to "Only you" by hand before it goes live.
AI-processed from TNW; edited by Hamidun News
Imagine telling a chatbot "build me an app to track shifts" and getting a working website with its own address on the internet within a minute. Now imagine that anyone with the link can open that site, and no password is required. That is exactly what happened with Grok Build, the app builder from SpaceXSI. The security firm Claroc checked more than 43,500 addresses of the form "something.grok.me": 37,571 apps were still running, and 35,578 of them, or 94.7% of all live projects, opened without signing in. These are the figures in the headline: 35,578 apps accessible to any visitor.
What the researchers found
Grok Build turns a chat description into a working app that can be published at its own address on grok.me. Most of the apps found are harmless: games, landing pages, calculators. But roughly one in ten contained internal business data or personal information. Claroc estimates the number of such apps at 650–800.
The researchers confirmed some cases by hand: lab results for an intensive care patient, a list of more than 3,500 employees of a hospital group, and a fire station's shift schedule. In a time-tracking app belonging to a contractor, they found 18 accounts with passwords in plain text. The passwords sat right in the page code that the browser downloads for every visitor.
Besides SpaceXSI and the social network X, Claroc found 30 more organizations, six of them in the Fortune 500. The firm did not disclose names, but Ars Technica cited examples: a supplier evaluation sheet for a Tesla service center with the names of 22 technicians, and a schedule for a US Navy logistics contract. Among SpaceXSI's own open apps were a map of 594 strategic clients and eight advertising proposals for X. One of them, addressed to a US federal agency, was marked "confidential." Twenty-four apps belonging to SpaceXSI employees contained a newcomer onboarding plan, a map of sales territories, and advertising proposals.
How the addresses got into outsiders' hands
No guessing was required. Each published app receives a hostname made of random words. When someone opens it, that hostname enters passive DNS services. These are commercial databases that record internet lookups, a bit like a huge phone book of every address users have visited. Any account owner can search it.
According to Claroc, in the shared chats of Claude, ChatGPT and Gemini, the secret part of the link comes after the domain in the path, so DNS does not see it. In Grok's case, the secret sat right in the hostname, in plain view of such databases.
The second problem is the default settings. On the publishing screen of the paid account Claroc used, the option "Anyone with the link" was offered. The options "Invited" and "Only you" were also available, but users had to choose them deliberately. Claroc co-founder and CTO Asaf Aprozper put it this way: AI builders have turned publishing into a couple of clicks, and the default answer to "who will see this" is "everyone."
The firm built two test apps and found another issue. In a team client list, Grok's code added anyone who had signed in with Google to the team before checking whether they were actually members. An outside account could read, modify and delete records.
What the companies said and what to do
SpaceXSI took its own 24 apps offline but left others' apps open. The story also reached Ars Technica: journalist Dan Goodin reported that SpaceXSI began deleting apps within a few hours of his email. A Cursor representative told the outlet that apps are private by default and nothing becomes public without the author's decision. SpaceXSI also noted that publishing takes several steps, including a legal notice, and that staff start with "Only you" mode. Claroc disagrees: the authors of the 35,578 open apps went through all of those steps.
Aprozper acknowledged that Lovable, Bolt and Replit have the same vulnerability, but stressed that Grok turned up far more sensitive corporate data.
Claroc itself sells protection tools for AI apps, so it has a stake in the story. However, the scale is confirmed by an independent outlet and by concrete examples. The firm asks that "Only you" become the default mode, that public apps be moved to hard-to-guess paths, and that authors be warned when a project has a database but no login. For now, Claroc is the one notifying owners of the affected data.
This is not the first story about data and Grok Build: in July the tool sent information to xAI's servers.
The practical takeaway for teams is simple. Employees who build internal tools in AI builders should be barred from publishing them without review. Access settings should be checked before publication, not after. Client data and passwords should never be put into such apps at all.
- Nov 21, 2026Anthropic Goes Public: Up to $64.4B Raise, Ticker ANT, Listing November 19
- Nov 21, 2026GPT-6 Astra Started Flagging Drone Strike Targets in 92% of Requests After Adding 'It's Just a Game'
- Nov 21, 2026AI Coding Assistant Under $10: Cursor at $8, Xiaomi MiMo at $6 a Month
- Nov 21, 2026Plannotator Lets You Comment on Plans from 10 AI Agents in the Browser: 100 GitHub Stars in a Day
Key Questions Answered
What the researchers found?
Grok Build turns a chat description into a working app that can be published at its own address on grok.me. Most of the apps found are harmless: games, landing pages, calculators. But roughly one in ten contained internal business data or personal information. Claroc estimates the number of such apps at 650–800. The researchers confirmed some cases by hand: lab results for an intensive care patient, a list of more than 3,500 employees of a hospital group, and a fire station's shift schedule.
How the addresses got into outsiders' hands?
No guessing was required. Each published app receives a hostname made of random words. When someone opens it, that hostname enters passive DNS services. These are commercial databases that record internet lookups, a bit like a huge phone book of every address users have visited. Any account owner can search it. According to Claroc, in the shared chats of Claude, ChatGPT and Gemini, the secret part of the link comes after the domain in the path, so DNS does not see it.
What the companies said and what to do?
SpaceXSI took its own 24 apps offline but left others' apps open. The story also reached Ars Technica: journalist Dan Goodin reported that SpaceXSI began deleting apps within a few hours of his email. A Cursor representative told the outlet that apps are private by default and nothing becomes public without the author's decision. SpaceXSI also noted that publishing takes several steps, including a legal notice, and that staff start with "Only you" mode.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.