A security firm says tens of thousands of small web apps that people built by chatting with Grok could be opened by anyone on the internet, with no sign-in. Some held staff rosters, customer lists and medical details, and 24 of them were built by employees of the company that makes Grok. This explainer covers what was found, why the links were easy to discover, what the company says, and how to check your own app.
The research was published on 10 November 2026. We checked again on 15 November and found no later announcement from the company, so details may still change.
What happened
Grok, the AI assistant from SpaceXSI (the merged SpaceX and xAI), has a Build mode. You describe an app in the chat, Grok writes it, and you can publish it to its own web address ending in grok.me.
On 10 November 2026 the security firm Claroc published a report on those published apps, and Ars Technica covered it the same day. Claroc’s main numbers:
| What Claroc counted | Number |
|---|---|
App names found under grok.me | 43,601 |
| Addresses it tried | 43,551 |
| Apps still online | 37,571 |
| Online apps that opened with no sign-in | 35,578 (94.7%) |
| Online apps that refused access (HTTP 401) | 1,993 (5.3%) |
Most of the open apps were harmless: games, landing pages and calculators. Claroc says about 1 in 10 held sensitive material and roughly 2% held material it rated high or critical. Its overall estimate is 650 to 800 apps with private data. These figures come from Claroc alone. The Cloud Security Alliance, which summarized the research, notes that it has not verified them independently. Claroc also sells security tools for AI-built apps, which is worth knowing as you read its conclusions.
A note on names: press coverage calls the product “Grok Build”. SpaceXSI’s documentation uses that name for a coding agent that runs in your terminal, and lists “Build mode” on grok.com separately. Claroc’s findings are about web apps published from the chat window, not code on your own computer.
What was exposed
Claroc says it only loaded the first page of each app once, without signing in, and did not name the outside organizations involved. Among the apps it confirmed by hand were:
- a hospital group’s staff event app with a searchable list of more than 3,500 employees
- an intensive care briefing for a single patient
- a fire department’s live shift board with names and badge numbers
- a staff timesheet whose code contained 18 accounts with passwords in plain text
It also found 24 open apps built by SpaceXSI and X staff, including an onboarding plan marked confidential, a sales territory map covering 594 accounts, and advertising proposals. Beyond those, Claroc identified 30 organizations with internal data in open apps. It says 26 of them have more than 1,000 employees and six are in the Fortune 500.
Why the links were easy to find
Grok suggests an address made of four random words, which nobody would guess. The problem is where the random part sits.
Each published app gets its own hostname, in the form <name>.grok.me. Every time someone opens that address, their computer asks a DNS server where the hostname lives. Some companies collect those lookups into searchable databases, a service known as passive DNS. So the first time the intended audience opens the link, the name can land in a commercial index, and one pattern search lists it alongside thousands of others.
Shared chats work differently. Claroc points out that Claude’s published artifacts and shared conversations on ChatGPT and Gemini put the random ID in the path, the part of the address after the slash. DNS never sees the path, so those links can’t be harvested this way.
| Kind of link | Where the secret sits | Visible to DNS? |
|---|---|---|
| Grok published app | Hostname (<name>.grok.me) | Yes |
| Shared chat on Claude, ChatGPT or Gemini | Path after the domain | No |
Two caveats. First, Grok is not alone: Claroc lists ChatGPT, Lovable, Replit, Base44, Google AI Studio and Meta’s Muse as builders that also give each app its own hostname. Second, a path is not automatically private. In August 2025, more than 370,000 shared Grok conversations turned up in Google search because the pages could be indexed.
The disputed default
A discoverable address is only a problem if the app behind it is open. This is where the two sides disagree.
- Claroc’s account: in the paid account it tested, the publish panel offered three choices, “Anyone with the link”, “Invited users” and “Only you”, and “Anyone with the link” was already selected.
- SpaceXSI’s account: a representative of Cursor, the coding company SpaceXSI owns, told Ars Technica that the apps “are private by default and nothing goes public unless the creator chooses to share it”. According to Claroc, the company also said publishing takes several steps and that employee accounts default to “Only you”.
Both sides agree that nothing goes public until the builder clicks Publish. They disagree about what the access setting shows at that moment for ordinary users. Ars Technica says the representative did not answer its follow-up questions.
For comparison, Claroc says an app it shared from ChatGPT was set to “Only me” until it changed the setting.
A second problem: what the apps let visitors do
Claroc also built two test apps with made-up data and published them with whatever settings Grok proposed.
The first, a shared contact list, had a database and no sign-in. A stranger could add a row that the next visitor saw. The second was described as a customer list for a four-person sales team. Grok added sign-in buttons on its own, but Claroc says the code added anyone who signed in to the team before checking membership, so a Google account that had never been invited could read, edit and delete records.
Claroc tested this only on its own apps, so it shows what Grok can produce, not that every open app is editable.
What SpaceXSI has done so far
According to both Claroc and Ars Technica, SpaceXSI took down the 24 apps built by its own staff. Claroc says the company had not replied to its report directly and did not say it would contact other app owners, so Claroc began notifying affected organizations itself.
Reports published on 11 November by the Cloud Security Alliance and heise online describe no change to the default sharing setting, and say other users’ apps were not taken down. Claroc says it proposed four fixes: serve apps on a hard-to-guess path under one host, add an “everyone in my organisation” option, make “Only you” the default, and warn at publish time when an app has a database but no sign-in.
How to check your own Grok app
If you have published anything from Grok’s Build mode, this takes a few minutes.
- Open your app’s address in a private or incognito browser window, where you aren’t signed in to Grok. If the app loads, anyone can open it.
- In Grok, find the app’s access setting. If the app holds anything internal or personal, switch it from “Anyone with the link” to “Invited users” or “Only you”.
- Repeat step 1. Claroc found that an app set to “Only you” answered a stranger with an error (HTTP 401) on both the page and its data.
- If nobody needs the app any more, unpublish it. In Claroc’s test, unpublishing took the address offline within seconds.
- Look through the app’s code for names, email addresses and passwords. Anything in a script file is downloaded by every visitor’s browser. Change any password that appeared there.
- If the app has its own sign-in, try it with an account from outside your team. If that account gets in, the app needs a proper list of who is allowed.
If you run IT or security for a team, the Cloud Security Alliance adds three steps: find out which AI app builders your staff use, test only apps you can tie to your own organization, and report apps with no identifiable owner to the vendor. If your own details were in an exposed app, see what to do after a data breach notice.
Why it matters
“AI app builders turned publishing into a few clicks, and the default answer to who can see it is everyone,” Claroc’s Asaf Aprozper says in the report. An app that would once have gone through an IT team, with company sign-in and an owner, can now be built and published from a chat window in minutes.
It is one of two Grok security stories this week, alongside the unrelated ArcticRAT Android malware that uses Grok. It also lands as SpaceXSI pushes Grok into office work with tools like the Grok Bot that joins your meetings. The more these tools do, the more their defaults matter.
Sources
Links last checked on 15 November 2026.
- Claroc: GhostShares 2, Grok apps exposed business secrets and private data
- Ars Technica: Grok Build coding agent leaks confidential info
- Cloud Security Alliance research note on exposed Grok apps
- heise online: Grok Build AI apps expose trade secrets and personal data
- SpaceXSI security FAQ (Grok Build CLI and grok.com Build mode)
- Fortune: shared Grok chats indexed by Google (August 2025)





Write a response