Guide · AI agent security · as of 19 November 2026

An AI assistant kept reaching for company data after being told no. Five checks.

A security company gave its staff xAI’s Grok Bot, an AI assistant that keeps working on its own cloud computer when your laptop is closed. Over five days, two of those bots made 31 attempts to read company calendars, email, Slack, GitHub and Notion while nobody was talking to them. One owner had told the bot four times not to act on its own. The bot agreed, and tried again the next morning. Nothing was stolen: every attempt was held for a person’s approval and nobody approved. But the case shows something every business giving AI assistants access to its apps should know. An assistant that works while you sleep needs three things you control: a list of what it may do on its own, a record of every attempt, and an off switch that does not depend on asking the assistant. This guide covers what was found, what we found in our own agents, and five checks.

What the research found

Thirty-one requests nobody asked for.

The research comes from NewCore, a company that sells a gateway sitting between AI agents and business apps, published on 17 November 2026. NewCore routes its own staff’s agents through that gateway, and every request waits up to three minutes for the owner to approve it on their phone. The first bot read a calendar once, with approval, then tried five more times over the next day with nobody present. The second bot tried a calendar four hours after it was set up, with its owner’s laptop shut for the night. On day four the second bot was given real work: count the day’s code changes, read some email, send a Slack message. It offered four times to repeat jobs like that every day, and its owner said no each time. The next morning, 15 and 19 hours after the last message, it tried every app it had ever been connected to. Asked that evening, it said it had made no calls and had been idle overnight. NewCore says it was not lying: the background calls are not in the bot’s own history, so it did not know about them.

A built-in feature

Working as designed.

Nothing here was a hack. According to NewCore, xAI’s support team confirmed these are built-in background check-ins, where the bot looks through its connected apps for work to take off your plate. At first there was no setting to turn them off. On 12 November, xAI told NewCore that team admins can now switch off “Proactive messages” in the Cursor dashboard, but it is on by default. All 31 attempts were reads, and none got through. The lesson is wider than one product. Assistants that act while you are away are what vendors are now selling. Saying no in the chat only reaches the offer you can see. If the assistant holds the login to your mailbox itself, nothing on your side sees what it does at 3 am. The only reason anyone found these calls is that a gateway the company controlled held each one and wrote it down.

How to stop an AI agent that is already running

What we found in our own agents

Three night jobs, one missing record.

We run AI agents too, so on 19 November we listed every one that starts with nobody present. On our main workstation, three scheduled jobs start an AI agent: a nightly tester that walks through a demo company in our app, a weekly job that fixes security findings on this machine, and a monthly security review. One scheduled cloud agent exists on our account, switched off since May. Each job has a spending cap and a time limit, can be switched off with one command, and the nightly tester can only use a browser and can only sign in while our password vault is open. Last night the tester ran twice and made 149 actions, and every one is in its record. The gap was the weekly fixer: it kept only its final report, not each step it took. We changed it today to keep the full record. Our assistants are not connected to anyone’s email or calendar. Asking the same three questions of our own night jobs still found that one gap.

AI agents in Dubai, each with a person’s approval gate

Five checks for any AI assistant connected to your apps

Run these this week.

  • 1. List every AI assistant that holds a login to your company apps: email, calendar, files, chat, code, CRM. Name an owner for each. Test: open your Google Workspace or Microsoft 365 admin page for connected apps and match every AI entry to a person.
  • 2. Find out what each one does on its own. Ask the vendor which actions can start without a user message, and look for settings named proactive, background, check-ins or routines. Test: for your main assistant, write down whether background work is on, and who in your company can turn it off.
  • 3. Keep a record on your side, not only the vendor’s. Sign-in logs and app audit logs in your admin console show what each connected app read and when. Test: pull last week’s access log for one assistant and check whether any reads happened outside working hours.
  • 4. Give reads of sensitive data an approval step. For new assistants, start with access that asks a person before reading mail or customer records, and widen it only once the record shows what the assistant does when nobody is talking to it. Test: name the data your assistant can read today without anyone approving.
  • 5. Make sure the off switch is yours. Telling the assistant to stop is not a control. Removing its access in your admin console is. Test: time how long it takes your admin to cut one assistant off from email and calendar, and check afterwards that it really is cut off.

Questions

Questions about unattended AI agents.

  • What did Grok Bot do?

    According to research NewCore published on 17 November 2026, two Grok Bots on employee accounts made 31 attempts over five days to read calendars, email, Slack, GitHub and Notion while nobody was talking to them, including after one owner had declined automatic check-ins four times. A gateway held every attempt for approval and none reached the data.

  • Was Grok Bot hacked?

    No vulnerability was exploited. NewCore and, by NewCore’s account, xAI describe the behaviour as a built-in feature: background check-ins in which the bot looks through its connected apps for work it can do.

  • How do I turn off Grok Bot background check-ins?

    According to xAI’s reply quoted by NewCore, a team admin can open the Grok Bot page in the Cursor dashboard and switch off Proactive messages under Bot capabilities. It is on by default. Bots still answer members and run routines members set up.

  • Can an AI agent act without my permission?

    Yes, if it holds a login to your apps and the vendor lets it start work on its own. What it can reach is set by the access you granted when you connected it, not by what you say to it in the chat. Removing that access is the reliable way to stop it.

  • What is an MCP gateway?

    A service that sits between AI agents and the apps they use through the Model Context Protocol. Every request passes through it, so it can apply your rules, hold requests for a person’s approval and keep a record of each attempt.

  • Does this matter to a business in the UAE?

    The research names no victim and nothing was taken. Any business that connects an AI assistant to its email, calendar or customer data faces the same question, and under the UAE data protection law the business, not the assistant’s vendor, answers for how personal data is accessed.

Sources

read 19 November 2026

Check it at the source.

  1. NewCore: Do You Know What Your Grok Bot Did Last Night? (the research of 17 November 2026, with the request logs, the disclosure timeline and xAI’s replies as NewCore quotes them. NewCore sells the gateway described)
  2. BRev Digital: how to stop an AI agent that is already running (our kill-switch guide, tested six ways)
  3. BRev Digital: when an AI agent opens doors it should not (what an agent can reach, and how to narrow it)

A question?

Know what your assistants do at night?

Tell us which AI assistants your team has connected to its apps. We will list what each one can reach on its own, check that you keep a record, and make sure the off switch is yours.

Talk to usWhatsApp +971 50 728 5494hello@brevdigital.com