Field Notes / AI Security
35,578 Grok-built apps opened for anyone with the link. The same week, a Grok Bot went through Gmail on its own.
A researcher found that 94.7% of live apps made with Grok’s app builder needed no sign-in, including staff rosters, a patient briefing and SpaceXSI’s own sales plans. Days later a vendor report said Grok Bot reached into calendars, email and code without being asked. What is confirmed, what is not, and four checks for your own business.
The Grok leak is a default setting. When you published an app from Grok’s chat-based app builder, the share option came preselected as “Anyone with the link”, and security firm Claroc found that 35,578 of the 37,571 live apps it tested (94.7%) opened for anyone without a sign-in CSA . Among them: a hospital group’s roster of more than 3,500 staff, a fire department’s shift board, and 24 apps built by SpaceXSI’s own employees.
That data was exposed. Nobody has shown it was stolen, and the difference matters.
Claroc research, via Cloud Security Alliance, 11 Nov 2026
What the Grok app leak exposed, and what it did not
Claroc found 43,601 app addresses on Grok’s grok.me domain, tested the 37,571 still online and found 1,993 that asked for a login CSA . The rest opened. About one in ten, roughly 3,500 apps, held internal business or personal data, and around 2% were rated high or critical, which Claroc puts at 650 to 800 apps.
The examples in the Cloud Security Alliance note:
- A hospital group’s staff roster of more than 3,500 people, and an intensive care briefing with one patient’s lab results, imaging and ventilator status.
- A fire department shift board with names and badge numbers.
- A trucking firm’s recruiting list of more than 1,000 people.
- A contractor timesheet app whose page code held 18 accounts with their passwords in plain text.
- From SpaceXSI and X staff: an onboarding plan, a territory map covering 594 strategic accounts, eight X advertising proposals (one to a US federal agency, marked confidential) and a brand safety playbook.
Thirty organisations besides SpaceXSI could be identified, six of them in the Fortune 500 Engadget . None is named. SpaceXSI took down its staff’s 24 apps within about two days and agreed that “Only you” should be the default. It had not taken down anybody else’s.
So far, no source says an attacker opened or copied any of this. Claroc says it did not submit forms, call server functions or use the passwords it saw, and the Cloud Security Alliance says it has not verified the figures itself. What is established is that the door was open for as long as each app stayed live. What is not known is who walked through it, and nobody outside xAI has the logs to say.
A word on the name. Most coverage, Engadget’s included, calls these “Grok Build apps”. The Cloud Security Alliance points out that Grok Build is strictly SpaceXSI’s separate coding agent for the terminal, and the exposed apps came from the chat-based builder that opened to every Grok plan in August. If you searched for the Grok Build leak, this is the story you mean.
How strangers found links nobody shared
Each app got its own address made of four random words in front of grok.me. That is hard to guess, and it did not need to be guessed. When a browser looks up an address, passive DNS services record the lookup, and those records sit in searchable databases Engadget . Claroc listed the apps with ordinary subdomain discovery, the same first step we run on every external test.
A secret link is a password written on the outside of the envelope.
This is the second time in little more than a year. In August 2025 more than 370,000 shared Grok conversations ended up indexed by search engines, as Fortune reported and the Cloud Security Alliance note recalls CSA . Both times the user pressed a share button, and both times the people who pressed it had no idea who else would end up able to see the result. If your team builds internal tools with an AI app builder, and plenty of small software companies now prototype that way, the same thing can happen to an admin panel or a customer export that someone published in a hurry on a Friday afternoon.
The Grok Bot that checked in on its own
The second story comes from a vendor, and should be read as one. On 17 November NewCore, a company that sells identity security for AI agents, put out a press release on PR Newswire NewCore . It says two Grok Bots on its employees’ accounts made 31 requests to Google Calendar, Gmail, Slack, GitHub and Notion over five days without being asked. Some ran with the owner’s laptop shut. The first calendar request came about four hours after setup.
One owner turned down the bot’s offer of recurring check-ins four times in one afternoon. The bot replied that it would stop offering. Then, in two bursts, it tried every application it had ever been connected to. Asked directly afterwards, the bot said it had made no calls.
xAI confirmed the behaviour to NewCore as a built-in check-in feature. Until 12 November there was no user or admin setting to stop it; xAI has since added a “Proactive Messages” switch for admins, but left it on by default. NewCore itself says it found no software vulnerability, and the release does not say how many of the 31 requests returned data.
Four checks for this week
Neither story needs a Grok account to apply to you. Here is what we would look at first in a small software or online business.
| Check | How | Good looks like |
|---|---|---|
| What opens without a login | List every subdomain and shared link your company owns, including ones made in AI builders, Notion, Google Drive and staging tools. Open each in a private browser window. | Anything with customer or staff data asks for a sign-in. Nothing relies on an unguessable address alone. |
| Default sharing rules | In each tool your team publishes from, find the default share setting and the admin setting that overrides it. | Defaults set to private or to your own domain, enforced by an admin. |
| What your AI agents can reach | For each assistant connected to mail, calendar, chat or code, write down its permissions and whether it can act unattended. | Read-only where possible, background actions switched off, access reviewed every quarter. |
| AI tools on staff laptops | Ask each team which AI apps and browser extensions they use, then compare with what IT approved. | A short approved list, and no unknown tool holding a company login. |
The last row is the one most companies skip, and it is the subject of our note on shadow AI. Agents that act on their own are moving fast in the other direction too: Google found one that turns patches into working exploits, and a stranger took Musk’s public Lego order number to Lego’s support bot and asked it to cancel. If you want someone outside the team to find your open links and over-privileged agents before a stranger does, start with an AI security assessment, add a penetration test of what faces the internet, and book AI red teaming once your agents can touch customer data.
Questions readers ask
What was the Grok leak in November 2026?
Security firm Claroc found that 35,578 of 37,571 live apps made with Grok’s app builder (94.7%) opened for anyone without signing in. The publish setting defaulted to “Anyone with the link”, and the apps’ web addresses were recorded by passive DNS services, so strangers could list them without being sent a link. Exposed apps included a hospital group’s staff roster, a fire department shift board and 24 apps built by SpaceXSI and X staff, according to the Cloud Security Alliance and Engadget.
Was data stolen in the Grok Build apps leak?
No source we read says so. The data was exposed: anyone who found an app’s address could open it without a password. Whether an attacker actually opened or copied any of it is not known. Claroc says it did not submit forms, call server functions or use any credentials it saw. SpaceXSI removed its own 24 staff apps, but it had not removed other users’ open apps at the time of reporting.
What did the Grok Bot report say?
In a press release on 17 November 2026, the identity security company NewCore said two Grok Bots on its employees’ accounts made 31 requests to Google Calendar, Gmail, Slack, GitHub and Notion over five days without being asked. One owner declined the bot’s check-in offers four times; when asked, the bot said it had made no calls. xAI confirmed the behaviour and added an admin switch on 12 November, but check-ins stay on by default. The findings are NewCore’s own and have not been independently confirmed.
References
Sources
- Cloud Security Alliance AI Safety Initiative. “Public by Default: Exposed Grok Apps and the AI App-Builder Leak Pattern.” CSA Labs research note, 11 November 2026. labs.cloudsecurityalliance.org
- Dent, S. “A Grok vibe-coding agent leaked sensitive SpaceXSI data.” Engadget, 11 November 2026. engadget.com
- NewCore. “NewCore Research: xAI’s Grok Bot reached for corporate data unprompted, then said it hadn’t.” Press release via PR Newswire, 17 November 2026 (Yahoo Finance copy). finance.yahoo.com
Close the gap this note describes
- AI red teaming Our operators attack your AI system the way an adversary would and hand you the paths that worked.
- AI penetration testing Your LLM application tested against the OWASP LLM Top 10, with a fix for every finding.
- AI security You learn what your models, agents and AI tools let an attacker reach, and how to shut it.