Grok Bot: what it is, what it can reach, and what to decide before you sign it in

Grok Bot is the always-on agent product from SpaceXSI, the company behind Grok, launched in beta on 11 August 2026 and at version 1.0 since 2 November. Each user gets a cloud computer with a browser, files and a terminal, and Bots work in the accounts you sign them into. Its docs say a Bot has no access by default and acts as you. So the risk is set by which accounts you connect and which actions need approval.

The team behind Keybrum's attacksUpdated 11 min read

In short

  • Grok Bot launched in beta on 11 August 2026 for SuperGrok and paid Cursor plans, opened to enterprises on 3 September and reached version 1.0 on 2 November. It runs only on cloud computers hosted by Cursor, in the United States.
  • Each user gets one microVM, separate from other users. All of that user's Bots share it: the docs say to treat a login or file on the computer as available to every Bot you run.
  • A Bot has no identity of its own. It acts as the signed-in member, so it can do what you can do in every account you sign it into.
  • Network allowlists, enforced Auto-review, audit logs and action recording are Enterprise only. Teams without a network policy default to allow-all.
  • The docs say their defenses against prompt injection reduce the risk and do not eliminate it. Keep sending, paying, deleting and deploying behind approval.
On this page

What Grok Bot is

Grok Bot is a product of AI agents, called Bots, that you message like a colleague and that keep working when you close the app. It was announced on 11 August 2026 by SpaceXSI, the company that publishes Grok at x.ai. The announcement describes it this way:

Bots have their own computer. They sign into the tools you already use and work across apps, inboxes, and more. They finish jobs end to end, and only come back when something needs your approval.
Introducing Grok Bot, SpaceXSI, 11 August 2026

At launch it was a beta for SuperGrok, SuperGrok Plus and SuperGrok Heavy subscribers, for Cursor Pro, Pro+ and Ultra, and for Cursor Teams Standard and Premium, on desktop and iOS. It has its own usage, separate from those plans. On 3 September 2026 it opened to enterprises, with what the company calls access, network and audit controls. Version 1.0 followed on 2 November 2026: Bots can be shared with a team, brought into Slack, Microsoft Teams and video calls, and can offer to take on work nobody asked for.

Three things set it apart from a chat assistant. A Bot works in a real computer, so it can use a website that has no API. You can run several Bots that message each other and hand work over. And a Bot can watch you do a job once, save it as a routine and repeat it on a schedule. Each of the three is also a security question, because the product is only useful once it holds your access.

What a Bot can reach

The security documentation, last updated on 13 November 2026, gives a clear picture. The desktop and mobile apps send each request to Cursor's servers. The work runs on a hosted computer. Prompts go to model providers, and tool calls reach other services through connected apps.

WhatWhat the docs sayWhat it means for you
The computerOne Firecracker microVM per user, with a browser, a filesystem and a terminal, hosted by Cursor in the United StatesOther users cannot reach it. It persists: files and browser sessions stay on a durable disk between sessions
Your BotsWithin one user, every Bot shares that computerA login given to one Bot is available to all your Bots
IdentityA Bot has no identity or credentials of its own and acts as the signed-in memberActions in your CRM or inbox are yours. A Bot cannot hold more access than you, and holds no less either once signed in
Connected appsOAuth tokens stay on Cursor's backend. The Bot calls tools without receiving the tokenA hijacked Bot cannot read the token, but it can still use the tool
PasswordsLogins saved in Vault and values typed in a Secure Form are filled into the page by reference. The model never receives themUse Vault or the Secure Form. Never type a password into the chat
Your own machineThrough the desktop app a Bot can run commands and read files locally. Per-command approval is the defaultThe docs recommend Never allow unless a Bot has a specific reason to work on local files
From the Grok Bot security page and security FAQ, November 2026.

The enterprise announcement sums up the default in one sentence: "A Bot has no access by default and reaches only the accounts you sign it into." That is a good default. It also means the size of the risk is set by you, one sign-in at a time.

What the vendor says about security and data

The documentation is more candid than most. It names its controls and, for several of them, their limits. Credit where due: per-user isolation in a microVM, tokens kept out of the model, and approvals shown with the proposed operation and its inputs are sound design choices.

  • Approvals and Auto Review. An independent review model evaluates risky actions before they run: shell commands, app calls, computer use, changes to routines, and delegation to other agents. It can let an action proceed, ask you, or deny it. The docs add that it does not review every side effect, and give memory writes and most settings changes as examples.
  • Network policy. Admins can restrict which destinations the computers reach. This is Enterprise only, and teams without a policy default to allow-all. The simpler blocked-sites mode is described as "best effort, not a security boundary": a request to a bare IP address gets through.
  • Logs. Audit logs are Enterprise only. Action Recording, which records what Bots did, is Enterprise only and off by default, with 90 days of retention when on.
  • Training and retention. With Privacy Mode enabled, customer data is not used for training, and model providers do not keep prompts or outputs. The docs note that providers may run abuse classifiers and store flagged data.
  • Models. Grok Bot chooses the model. There is no model picker, and a team's model allowlist does not govern it.
  • Hosting and certifications. It runs only on Cursor-hosted computers, with no on-premises option. Anysphere, the company behind Cursor, holds ISO/IEC 27001 and ISO/IEC 42001 certifications that include Grok Bot in scope.

On prompt injection, the docs say that outside content is marked as untrusted when shown to the model, and that Auto Review checks actions against your request. Then they state the limit:

These controls reduce, but do not eliminate, risk from malicious content, which is another reason to keep consequential actions behind approval.
Grok Bot security, SpaceXSI Docs

The risks that follow from that access

None of the points below is a flaw in Grok Bot. They are what any agent with a browser, a terminal and your sessions makes possible. The OWASP list for agentic applications names each of them.

  • A page or an email can steer the Bot. A Bot that reads your inbox and can also send mail reads text written by strangers with the right to act on it. The vendor says its defenses reduce this risk without removing it.
  • One shared computer per user. A Bot you created for a small job can use the session another Bot opened in your bank, CRM or cloud console. The FAQ suggests a separate user when a workload needs its own computer and credentials.
  • Sessions outlive the task. The docs say that deleting a Bot does not remove computer files or browser sessions. Signing out and revoking is your job.
  • Routines run without you. A saved routine repeats a job on a schedule. A mistake in it repeats too, and changes to routines are one of the things an attacker would want a hijacked Bot to make.
  • Payments. A card saved in Vault fills at checkout without a separate approval each time. The docs say that members who want to approve each payment should not save the card.
  • Shared and proactive Bots widen the audience. Since version 1.0 a Team Bot works with several people and in Slack or Microsoft Teams channels, and a Bot can start work on its own initiative. The vendor says Bots respect each person's permissions and ask before sensitive actions. More people and more channels still mean more text the Bot reads that you did not write.
  • Trust grows faster than checks. The launch post quotes an employee who lets a routine run "without me verifying and reviewing". OWASP calls this human-agent trust exploitation: people approve, or stop looking, because the agent has been right so far.

What to decide before you give it credentials

  1. 1

    Choose the accounts, one by one

    Start with accounts where a mistake is cheap: a calendar, a drafts folder, a test workspace. Leave out the bank, the domain registrar, the cloud console and the production database until you have a reason and a limit.

  2. 2

    Prefer a restricted account to your admin one

    The Bot acts as whoever is signed in. Where a tool lets you create a member with fewer rights, sign the Bot into that one. For code hosts and databases, a read-only token or role.

  3. 3

    Write Ask first rules

    In Settings, General, Auto-review, add rules under Custom Rules. The docs give the model: "ask first before sending any external email". Add the same for payments, deletions, permission changes and deploys. Ask first wins when rules conflict.

  4. 4

    Turn local execution off

    In Settings, Computer, Your computer, set Execution on this computer to Never allow, unless a Bot needs your local files for a job you chose.

  5. 5

    State the stop in the request

    The docs call the request itself the strongest boundary. Say what the Bot may change and where it must stop: "draft the budget change, do not change the campaign or message the agency".

  6. 6

    Plan the clean-up

    When a project ends: pause or delete its routines, sign out of the websites on the computer, remove the connected apps and revoke their authorization in the source service, and remove sensitive files from /workspace.

For a company, add one rule from the docs: an approval covers the proposed action, and nobody should approve an action whose target or effect they cannot identify. If a Bot was installed by an employee without anyone knowing, find it before it is signed into a shared account.

What to check on anything a Bot builds or deploys

The announcements show engineering Bots that reproduce bugs, fix them, watch pull requests and keep builds moving. A Bot with a terminal can also write a small web app and put it online. The result is ordinary generated code, and it fails in the ordinary ways. The agent that wrote it was told to make it work. Nobody told it who must be refused.

  • Open the app in a private window, signed out: no page or API route should return another person's data
  • Create two test accounts and check that the second cannot read or change the first one's records
  • Search the published JavaScript for secret keys: sk_, service_role, private keys, provider tokens
  • If the app uses Supabase or Firebase, read the access rules of every table and bucket
  • Check that AI routes and webhooks refuse a caller who is not signed in or not signed
  • Check that no .env, backup or .git folder is served
bash
# On your own app only. Replace the domain.
# 1. Is a .env file served? Expect 404.
curl -s -o /dev/null -w "%{http_code}\n" https://your-app.example/.env

# 2. Which protective headers are set?
curl -sI https://your-app.example | grep -i -E "content-security-policy|strict-transport-security|x-frame-options"

# 3. Does an API route answer with no session? Use a route from your own code. Expect 401.
curl -s -o /dev/null -w "%{http_code}\n" https://your-app.example/api/orders
Three quick checks on an app a Bot shipped. A 200 on the first or the third needs a fix before launch.

Deeper guides: the vibe coding security checklist for the whole list, Supabase RLS for database rules, and API keys in the frontend for the bundle.

Where Keybrum fits, and where it does not

Keybrum is the automated pentest we build for web apps made with AI. It does not test Grok Bot. The Bots run on computers hosted by Cursor, and only the vendor can test those. It will not tell you whether your Auto-review rules are right or whether a routine is safe.

It tests the thing a stranger can reach: the app. For an owner who has proved ownership, it attacks the published app signed out and as two accounts, to find tables and API routes that return other people's data. It looks for secret keys in the JavaScript, a served .env, AI routes that answer anyone and Stripe webhooks that accept unsigned events. If the app runs on your own server, it checks that server for databases, admin consoles and AI tools such as Ollama or n8n open to the internet. Each finding comes with the request that worked and a fix prompt for your coding agent.

The one-time report costs €49 for one app. Plans start at €29 a month, excluding VAT, and attack again on every deploy, which fits a workflow where a Bot ships changes overnight.

Questions

What is Grok Bot?

Grok Bot is a product of always-on AI agents from SpaceXSI, the company behind Grok, launched in beta on 11 August 2026 and at version 1.0 since 2 November 2026. Each user gets a cloud computer, and Bots sign into your tools to finish jobs such as inbox triage, sales follow-ups or bug fixes. You message them from the desktop or mobile app, and since 1.0 in Slack and Microsoft Teams.

Is Grok Bot safe to use?

Its documented design is careful: one isolated microVM per user, connector tokens kept out of the model, and approvals for risky actions. The risk depends on what you sign it into. The vendor states that its defenses against malicious content reduce the risk and do not eliminate it, so keep consequential actions behind approval.

Can Grok Bot see my passwords?

According to the docs, a login saved in Vault or typed in a Secure Form is filled into the page by reference and the model never receives it. The Bot can see which logins are saved. In chats in the app it can also see the number and security code of a card saved in Vault, so that it can type them at checkout.

Is my data used to train models?

The security page says that with Privacy Mode enabled, customer data is not used for training and model providers do not retain prompts or outputs. A team's privacy mode governs its members. Check the setting on your own account, because the page describes the team case.

Where does Grok Bot run?

Only on cloud computers hosted by Cursor, in the United States. On-premises deployment and bring-your-own-image are not supported. A US-only data residency program exists for enrolled customers.

Does Keybrum test Grok Bot?

No. Grok Bot is a hosted product and only its vendor can test it. Keybrum attacks the web app you own, including one a Bot wrote or deployed, and the server behind it.

Sources

  1. 1.Introducing Grok Bot, SpaceXSI
  2. 2.Introducing Grok Bot 1.0, SpaceXSI
  3. 3.Grok Bot for Enterprise, SpaceXSI
  4. 4.Grok Bot security, SpaceXSI Docs
  5. 5.Grok Bot security FAQ, SpaceXSI Docs
  6. 6.OWASP Top 10 for Agentic Applications for 2026, OWASP GenAI Security Project

Your app is public. So are its mistakes.

Check what any attacker can already see in it, free, in ten seconds. The full attack, with proof and fixes: €49 once, or €29 a month on a plan.

or continue with Google or GitHub