Grok Bot became available to enterprises on September 3, 2026, and existing Grok and Cursor Enterprise customers got two weeks of free usage, with the option to invite their whole organization. Whether a company came in through that offer, an individual plan, or a self-serve Teams plan, someone there may already be using it. That’s not a guess. It’s how fast moving teams behave with an always on AI agent that plugs into tools they already use.
Why This Is a Leadership Problem, Not Just an IT Problem
Here’s the thing about a persistent AI agent versus the chatbot most people picture. A Grok Bot doesn’t wait for you to open a tab and ask it something. It sits inside a dedicated cloud computer, connected to whatever accounts a team member signs it into, running research, drafting, browsing and automations on its own schedule. For a sales team, that’s a feature. For a finance team, a law practice, or a private equity firm with cap table data, LP reporting exports and diligence materials in the same tools the Bot can reach, it’s a different conversation.
Most of that conversation happens after the tool is already live, not before. That’s backwards for anything sensitive. A governance checklist for this kind of rollout has to run before onboarding, not as a postmortem after something lands in the wrong place.
What a Grok Bot Is, and Isn't
Worth being precise here, because the marketing framing and the architecture aren’t quite the same thing.
Grok Bot runs on Cursor’s infrastructure (Anysphere, the company behind Cursor, operates it). Each user gets a dedicated Firecracker microVM, a hardware isolated computer, and every Bot that user creates shares that one computer. A Bot has no identity or credentials of its own. It can never hold more access than the person it’s signed in under, and connector tokens stay on Cursor’s backend rather than living inside the Bot’s session. One exception worth knowing: team managed connectors may use team or service account credentials.
That’s a different risk model than an autonomous agent with its own root access. The real risk isn’t the Bot exceeding a person’s access. It’s a Bot quietly using the full breadth of whatever that person can already reach, with nobody watching closely enough to notice.
So you’re not evaluating whether Grok Bot can break out of a sandbox. You’re evaluating whether your company has set the boundaries around what a signed in team member, and therefore their Bot, can touch.
1. Confirm Which Plan You're On
Start here, because it changes everything below it. Grok Bot ships on three tiers: Individuals, Teams and Enterprise. The self-serve Teams plan and individual plans do not include most of the controls that matter for sensitive data. Network Controls, the organization wide enable switch, Team Setup, Action Recording, computer management, audit logs, OpenTelemetry Export, SCIM provisioning and the MCP allowlist are all Enterprise only. A self-serve team can’t set a destination allowlist. Teams without a network policy default to allow-all.
So the first item on this checklist is a question: is this company on Enterprise, or did someone start using Grok Bot on an individual or self-serve Teams plan? If it’s the latter, the controls in steps two through eight mostly don’t exist yet. The sensible move is pausing the rollout until it’s properly licensed, not layering governance language onto a plan that can’t enforce it.
2. Set the Network Policy Before Anyone Signs In
Once you’re on Enterprise, Network Controls offers three modes: Allow All Network Access, Defaults + Team Allowlist, and Team Allowlist Only. Teams that never set a policy get allow-all. For anything touching sensitive data, Team Allowlist Only is the right starting position, expanded on purpose rather than opened broadly and narrowed later.
A few details to know going in. Running computers pick up a policy change within about a minute, and sleeping computers apply it when they next wake, so you don’t need to recreate them. Egress runs through shared static IP addresses across every Grok Bot customer, not a dedicated range for your company, and dedicated IPs aren’t available. That matters if a data room, banking portal or counterparty system restricts access by source IP.
3. Turn On Audit Logs and Action Recording, Separately
This is the one most rollouts get wrong, and it’s easy to miss because the two settings sound like they’d cover the same ground.
Audit logs cover admin, security, authentication and Grok Bot control plane events, viewable in the dashboard or streamed to your SIEM. Action Recording is a separate pipeline that captures what a Bot did, including scrubbed shell commands, and it’s off by default even on Enterprise. Audit logs tell you who signed in and when. They tell you nothing about what a Bot did with a signed in session. For that you need Action Recording switched on deliberately, and if you want those events in your own collector rather than Cursor’s internal 90 day store, OpenTelemetry Export has to be configured on top of it, also a separate step.
The practical instruction: enable both, don’t assume one implies the other, and confirm someone can produce an Action Recording export on request, not just say the setting is on.
4. Turn On Auto Review Enforcement and Write Rules That Mean Something
Auto Review is an independent model that evaluates risky Bot actions (shell commands, plugin calls, computer use, automation writes and delegation to other agents) before they run. Members can switch Auto-review off for themselves unless an Enterprise admin turns on Enforce Auto-review, which is off by default. Turn it on before you rely on team rules.
On Enterprise, add team Auto-review rules. Grok Bot’s recommended baseline puts production deployments, external email, payments and accepting legal terms behind Ask first, and Team Rules such as never moving company data to personal accounts apply to every member’s Bots. Put your own never events there too: no sending cap table, client or LP data to external email, no production changes, no accepting terms or completing payments through the Bot.
5. Give Sensitive Work Its Own Cursor User
Here’s a detail that’s easy to skim past. Every Bot one person creates shares that person’s single computer. A login, a browser session or a file in that environment is available to every Bot that person runs, not just the one that opened it. If someone runs a general research Bot and a Bot tied to a sensitive project from the same account, both effectively share one environment.
When a workload needs its own computer and credential set, give it its own Cursor user rather than a second Bot under the same person. It costs a seat. It’s worth it compared to finding out later that a routine research task had standing access to a live data room because they happened to share a computer.
6. Know Whose Certification You're Relying On
If a due diligence or vendor questionnaire asks about Grok Bot’s compliance posture, get the attribution right. The ISO/IEC 27001 and ISO/IEC 42001 certifications covering Grok Bot belong to Anysphere, the company behind Cursor, issued by Schellman, with Grok Bot in the current scope. Pull the certificates from Cursor’s trust portal (trust.cursor.com) rather than repeating a vendor’s marketing language into your own memo. A file that credits the wrong entity is the kind of error a client’s counsel or an acquirer will eventually notice.
7. Check Data Residency Against Your Obligations
Grok Bot computers run in the United States today. On premises deployment, deployment inside your own perimeter and bring your own image deployment are all unsupported, and Cursor’s US only residency program doesn’t apply to Grok Bot by default. If your company has European operations, foreign investors or clients with data location terms, that needs to sit in the same review as the network and audit questions, not get treated as a separate legal matter that never reconnects with the security checklist.
8. Put Computer Termination in Someone's Job Description
When a team member leaves, or a project ends and access should too, an organization admin can terminate that person’s computer from the dashboard, an Enterprise only control. The durable disk persists and a fresh computer starts on the next session, so this is a session cutoff, not a data wipe. It needs pairing with revoking the person in your identity provider. Neither step alone finishes the job.
Name who owns this in the offboarding checklist before the first Bot goes live, not after the first departure raises the question.
What This Checklist Doesn't Cover
Worth saying plainly: none of this stops a well crafted prompt injection from steering a Bot through content it reads on the open web, and Auto Review doesn’t catch every side effect. Memory writes and most settings changes fall outside it. Model selection sits with Cursor too. There’s a team model allowlist, but enforcement isn’t guaranteed, and onboarding includes an explicit acknowledgment that Grok Bot may not follow it. Layered controls reduce risk here. They don’t eliminate it, and a memo that implies otherwise is overselling what you can configure.
Where This Fits in a Broader Review
Eight settings, most of them Enterprise only, are a start, not a finish line. They’re also the kind of thing that gets configured once at rollout and never revisited as a company grows, adds teams or takes on new clients with their own requirements. Unapproved AI tools are the wider version of the same problem. Our 30-day plan for shadow AI walks through inventory, policy, and approved tools.
It’s the same lens we bring to an AI agent due diligence review for PE and M&A: what an agent can reach today, not what a settings page implies it can reach. It pairs with the containment principles in our CISO’s guide to agentic AI, and with the identity question underneath all of it, covered in our piece on permission creep and cloud permissions management, since a Bot’s access is only ever as tight as the human account it rides on.
If your company is about to roll out Grok Bot or another AI agent, or someone started using one without asking first, an AI Readiness Snapshot is the right next step before a Bot gets near sensitive data.
For the technical detail from the source, Grok Bot’s security documentation and the teams and enterprises guide are both worth reading end to end before your next AI tool decision, not after.




