Grok Bot only runs commands on your own desktop if you allow it in Settings → Computer, and the default is to ask you each time. Keep that default, add "Ask first" rules for sending, publishing, deleting and purchasing, and be careful with "Always allow". The setting most people miss is not a toggle at all: all of a user's Bots share one cloud computer, with the same files, browser sessions and logins, so splitting work across Bots does not separate it. This guide is based on xAI's and Cursor's docs as of November 2026.
Where does Grok Bot actually run?
Grok Bot can work in two places:
- Its cloud computer, a Firecracker microVM that xAI runs for you, with its own browser, files and logged-in sessions.
- Your own desktop, if you install the desktop app and allow local commands.
These carry very different risks. A mistake in the cloud computer affects that machine and whatever you have logged into there. A mistake on your desktop can touch your personal files, apps and anything your user account can reach.
What does "Ask every time" protect?
xAI's FAQ says local desktop commands run only if you allow them in Settings → Computer, and by default the Bot asks you each time. That gives you a checkpoint before every command on your machine.
The buttons, as reported
According to Digital Applied's launch coverage, the desktop prompt offers "Allow once", "Deny" and "Always allow", while the iPhone app offers "Approve once" and "Deny". The same report says Require Approval rules override Always Allow. These labels come from secondary coverage and may differ in your version of the app.
| Choice | What it does | When to use |
|---|---|---|
| Allow once / Approve once | Runs this one command | Default for anything you have not seen before |
| Deny | Blocks it | Anything unexpected or unclear |
| Always allow | Stops asking for matching commands | Rarely; only trusted, reversible, local-only actions |
Button labels per secondary coverage; default behaviour per xAI's FAQ, as of November 2026.
Read why "Always allow" is the riskiest button before you press it on your own computer.
Approvals expire
For tasks started without you, such as scheduled work, approval requests expire after about 30 minutes. An expired request is not approved. If you miss one, check the task and restart it.
Why does the shared computer matter?
This is the point xAI is most direct about: all of a user's Bots share one cloud computer, including files, browser sessions and logins. Its docs say: "Do not use separate Bots as a security boundary."
In practice:
- If you sign into your bank's site for one Bot's task, another Bot on the same account may be able to use that session.
- A file one Bot downloads is visible to the others.
- A Bot that reads untrusted web content runs on the same machine as your logged-in sessions.
- Bot A logs into a site
- Session saved on shared computer
- Bot B reads a hostile page
- Bot B can reach Bot A's session
Shared or team Bots get their own computer, according to xAI. If you need genuine separation for a sensitive task, that is the documented route. Otherwise, sign out of sensitive sites on the cloud computer when a task ends.
What "Ask first" rules should I set?
xAI says "Ask first" rules cover sending, publishing, deleting and purchasing. Keep all four on.
- ComputerLeave Settings, Computer on asking every time.
- SendingAsk first before any email or message to other people.
- PublishingAsk first before posting anywhere public, including X.
- DeletingAsk first, and keep bulk deletion manual.
- PurchasingAsk first; payments go through takeover anyway.
Each Bot also has its own email address at @mail.grokbot.com. Unrequested outbound mail from that address waits as a draft rather than sending.
What does Auto Review do?
xAI describes Auto Review as a check on planned actions, but says it "should complement, not replace, least privilege and explicit approval boundaries", and that safeguards "reduce, but do not eliminate, risk from malicious content". In other words, xAI tells you not to rely on it alone. The FAQ does not otherwise discuss prompt injection. See indirect prompt injection examples for what that risk looks like.
How are passwords and payments handled?
Passwords, 2FA, CAPTCHAs and payments go through a takeover flow, where you act directly. Credentials go via Vault, a password manager or a Secure Form. Never paste a password into a Bot's chat. Our guide to agent takeover, passwords and 2FA covers the safe way to do it.
What about work accounts?
On enterprise plans, Bots have no access by default and admins get audit controls. If you use Grok Bot through work, your employer's settings apply on top of yours. See personal AI agents at work.
How do I check what a Bot did?
Usage is at grok.com/dashboard → Usage. For a fuller review of actions, see our guide to auditing your agent's activity log. To remove a Bot's access to Google or Microsoft accounts, see revoking AI agent access.
Should I let Grok Bot use my own computer at all?
Only if a task genuinely needs local files or apps that the cloud computer cannot reach. Most research, drafting and web tasks work in the cloud computer, which keeps mistakes away from your personal files. If you do enable local access, use a separate, non-admin user account on your computer where possible, and keep "Ask every time" on. The joint government guidance Careful adoption of agentic AI services recommends never granting agents "broad or unrestricted access, especially to sensitive data or critical systems". It was written for organisations, but your own laptop is the closest thing you have to a critical system.
Bottom line
Leave computer access on "Ask every time", keep "Ask first" rules on for sending, publishing, deleting and purchasing, and treat the shared cloud computer as one environment. Sign out of sensitive sites after use, and use a shared or team Bot if you need real separation. For setup, see our sister site's Grok Bot setup guide and its Grok Bot review.



