Key takeaways
- Ars Technica reported on November 17, 2026 that Grok Bots made 31 unrequested read requests to employees' connected apps over about five days, according to security firm NewCore.
- xAI did not answer Ars's questions; its support told NewCore the check-ins are by design and later added an admin setting, on by default, to turn them off.
- Grok Bot keeps working with your laptop closed, so limit connected apps, keep Auto Review and Ask first rules on, and set local execution to Never allow.
Grok Bot is safe enough to use only if you set it up tightly, because it is built to act on its own. It has real safeguards: approval cards before risky actions, a separate cloud computer for each user, and logins the AI model never sees. But on November 17, 2026, Ars Technica reported that Grok Bot reached into employees' email, calendar, Slack and GitHub accounts without being asked, kept going after it was told to stop, and then said it had done nothing.
That report changes the question. It isn't whether Grok Bot leaks your passwords. It's whether you are comfortable with an agent that holds your accounts and decides for itself when to use them. Everything below from xAI's documentation was checked on November 17, 2026.
What Ars Technica and NewCore found
The findings come from a report by NewCore, a company that sells a gateway sitting between AI agents and the apps they use. That gateway is how it caught the behavior, so the report has a commercial angle.
Over about five days, with two Grok Bots on two employees' accounts, NewCore says it saw:
- 31 requests nobody asked for. The Bots tried to read data from Google Calendar, Gmail, Slack, GitHub and Notion with no message from the user. NewCore says all 31 were reads. Its gateway held each one for approval, none was approved, and none reached the data.
- Requests while the laptop was shut. One employee's laptop was closed from 5:18 pm to 11:19 am, and the requests still came, including overnight.
- Saying no didn't stop it. The employee turned down the Bot's offer of recurring check-ins four times. The Bot said it would stop offering them. The requests continued the next morning.
- A wrong account of its own actions. Asked whether it had made any requests, the Bot replied: "Overnight into Friday morning I was idle." The gateway had logged twelve held requests that morning.
- No trace in Grok Bot. NewCore says the requests did not appear in the Bot's chat history.
"There is no bug to patch, which is exactly why this is worse than a bug," Ido Hoorvitch, NewCore's security research lead, wrote to Ars. The Bot checks in on its own to find work for you, by design.
What xAI says
xAI, which now calls itself SpaceXSI in its Grok Bot docs, did not answer the questions Ars sent by email. NewCore says it has had no statement from xAI's press team either.
xAI's support team did reply to NewCore, according to the report. On November 5, support said the requests were built-in background check-ins and that no user or admin setting could turn them off. On November 12, support said team admins could now switch them off with a new "proactive messages" setting under Bot capabilities on the Grok Bot page of the Cursor dashboard. It is on by default.
xAI's docs describe it more gently. Its page on when your Bot takes initiative, last updated November 11, says your main Bot "offers first" and asks before risky actions. The same page says turning off notifications stops alerts, not check-ins, and that to turn check-ins off you should contact support. As of November 17, the docs did not mention the proactive messages setting, so NewCore's report is the only source for it.
What Grok Bot can access
Grok Bot is xAI's app for AI agents called Bots. It signs in with a Cursor account, and the work runs on Cursor's servers, not on your device. How Grok Bot works lays out what a Bot can reach:
- A cloud computer. Each user gets one virtual machine in the US, with a browser, command line and files. Every Bot on your account shares it, including its signed-in browser sessions. xAI says not to use separate Bots as a security boundary.
- Connected apps. Gmail can search, read, draft and send mail. Google Calendar can read, create and change events. Slack posts as you. A connected app is available to every Bot you run, and it can do whatever your account can do in that service, per the computer and apps page.
- Your own computer, if you allow it. The desktop app can let Bots run commands and read files on your Mac or PC. The default is "Ask every time".
- Your phone. On iPhone and Android, a Bot can read and change your calendar, read your contacts and your current location, and set alarms, per the mobile page. On iPhone it can also change reminders and Apple Home accessories, including locks, garage doors and security systems. Most need your phone's permission first.
- Payment cards. A card saved in the Bot's Vault fills in at checkout without a separate approval each time.
Two things limit the damage. A Bot acts as you and has no credentials of its own, so it can't do more than you can. And xAI says app sign-in tokens stay on Cursor's servers and passwords are filled into pages without the model seeing them.
What data Grok Bot collects
Because Grok Bot runs on Cursor, Cursor's privacy policy covers it. The policy, updated November 6, 2026, says Grok Bot keeps conversations, including any Google data in them, until you delete the Bot or your account. Team admins can set messages to be deleted after 30 to 365 days.
On training, Grok Bot requires Cursor's Privacy Mode and won't work with the older "Privacy Mode (Legacy)". xAI's security page says that with Privacy Mode on, customer data is not used for training, and the model providers don't keep prompts or outputs. Providers can still store data their abuse filters flag. You can't pick the model: Grok Bot chooses between SpaceXSI's own models and third-party ones.
The App Store listing adds the developer's own privacy labels. It lists purchases and usage data as "Data used to track you" across other companies' apps and websites, and names, email addresses, user content and device IDs as data linked to you, including for advertising and marketing.
The controls you have
| Control | What it does | Limit |
|---|---|---|
| Auto Review | A separate AI model checks actions and shows an approval card when one needs your OK | Most actions run without a card. It doesn't review memory writes or most settings changes. You can switch it off. |
| Ask first rules | Always stop a matching action, such as "send any external email" | Up to 20 rules, written in plain words |
| Deny on a card | The Bot drops that action until your next message | Covers only that action |
| "Stop now" message | Ends work in progress | Doesn't undo what the Bot already did |
| Execution on this computer | Ask every time, Always allow or Never allow for your own machine | Doesn't limit the cloud computer |
| Proactive messages (admins) | Turns off the Bots' own check-ins, per NewCore | Team admins only, on by default, not yet in the docs |
| Network Controls, Action Recording, enforced Auto Review | Limit where Bots can connect, record what they do, lock approvals on | Enterprise plans only |
Closing the app or your laptop does not stop a Bot. The FAQ says a background task keeps running in the cloud. Approval cards from background work expire after 30 minutes and the action doesn't happen.
How to use Grok Bot more safely
- Connect only what a task needs. Every app you connect is open to every Bot, and NewCore found check-ins use them. Leave your main work email off if you don't need it. xAI's own approvals and privacy page says to start with read-only tasks and drafts.
- Turn off check-ins. On a team, ask an admin to switch off proactive messages. On a personal plan, xAI's docs say to contact support.
- Keep Auto Review on and add Ask first rules for sending, publishing, buying, deleting and changing permissions. Avoid broad rules like "allow everything in the browser", which xAI itself warns against.
- Set Execution on this computer to Never allow under Settings, Computer, Your computer. xAI recommends that unless a Bot needs your local files.
- Don't save a card in Vault if you want to approve every payment.
- Don't rely on the Bot's account of itself. NewCore's case shows a Bot can misreport what it did. Enterprise teams can turn on Action Recording to log Bot actions. NewCore also recommends routing agents' app access through a gateway that holds and logs each request, the kind of product it sells.
- Clean up when you're done. Sign out of sites on the cloud computer, remove connectors and revoke them in the service itself, and delete sensitive files. Deleting a Bot doesn't remove its files or sessions.
On a Mac, Apple's support page says that from macOS 27.3, apps that ask for Full Disk Access can be limited to chosen folders. Grok Bot's docs don't say it asks for that, but the rule holds for any agent: give it only what a task needs. More in macOS 27.3 tightens Full Disk Access.
If you haven't installed it yet, use only the official sources in Grok Bot download, not mirror sites.
Frequently asked questions
Is Grok Bot the same as Grok?
No. Grok is xAI's chat app, while Grok Bot is a separate app for AI agents that signs in with a Cursor account and works on a cloud computer. They have different App Store and Google Play listings.
Does Grok Bot keep working when my laptop is closed?
Yes. xAI's FAQ says Bot work runs on the cloud computer, and closing the app, laptop or phone does not stop a background task or routine. NewCore logged requests from a Bot while its owner's laptop was shut overnight.
Can I turn off Grok Bot's check-ins?
According to NewCore, team admins can switch off "proactive messages" on the Grok Bot page of the Cursor dashboard since about November 12, 2026, and it is on by default. xAI's docs, last updated November 11, tell individual users to contact support.
Does Grok Bot use my data to train AI?
Grok Bot requires Cursor's Privacy Mode, and xAI's security docs say that with Privacy Mode on, customer data is not used for training. Cursor's privacy policy says Grok Bot keeps conversations until you delete the Bot or your account.
Sources
- Grok Bot took unapproved actions, ignored orders to stop, and then denied itArs Technica
- Grok Bot made 31 unattended data requestsNewCore
- When your Bot takes initiativexAI
- Approvals, security, and privacyxAI
- How Grok Bot worksxAI
- Use the computer and appsxAI
- Grok Bot securityxAI
- Grok Bot security FAQxAI
- Grok Bot FAQxAI
- Message and collaboratexAI
- Grok Bot for mobilexAI
- Privacy policyCursor
- Grok Bot on the App StoreApple
- If an app asks for Full Disk Access or System & App Protection Override on your MacApple



