Grok Bot shared-computer security audit

What does this setup do?

Before you give Grok Bots more connectors or routines, run a read-only audit of what every Bot on your account can see on the shared cloud computer. Connected accounts, files, approval rules, and stale access, then fix issues yourself after review.

Who is it for?

Anyone running multiple Grok Bots on one account who treats separate Bots as “isolated” without checking the shared browser, `/workspace` files, and connector list.

How do you set it up?

Tools

• A dedicated auditor Bot (or a fresh chat) with read-only instructions. Do not use the same Bot you are auditing as the only reviewer
• Grok Bot desktop or web to open Computer, connectors, and approval settings
• Optional: spreadsheet to track findings and revokes

Steps

1. Read xAI’s shared-computer docs once: one cloud computer per user, not per Bot; separate Bots are not a security boundary.
2. Create or pick an auditor Bot whose only job is this report. Do not grant it send, spend, or publish permissions beyond what you need to inspect settings.
3. Paste the audit prompt below. Ask it to list findings only. No revokes, deletes, or setting changes until you approve each item.
4. Walk the six sections: connected accounts, sensitive info on disk, agent/Bot access, approval rules, activity you can see in run history, and unused connectors or routines.
5. You sign out of sites, revoke connectors, delete files, and tighten Ask-first rules in Grok Bot settings. Never paste passwords or API keys into the chat transcript.
6. Re-run monthly or after adding a marketplace template, webhook routine, or new connector.

SHARED CLOUD COMPUTER SECURITY AUDIT (read-only report)

You are helping me audit my Grok Bot account's shared cloud computer. xAI documents that all Bots on my user share one computer: browser sessions, files under /workspace, and command-line credentials are shared, not isolated per Bot.

Rules for this run:
- Report only. Do not revoke connectors, delete files, change approval rules, send mail, spend money, or publish anything.
- Do not ask me to paste passwords, API keys, recovery codes, or webhook secrets into chat. If you need a credential location, describe the file path or setting name only.
- Do not rely on a Bot that is being audited to audit itself without a separate auditor perspective. I will review your output.

Inspect what you can see and return markdown with these six sections:

1) CONNECTED ACCOUNTS: List connectors/plugins and signed-in browser sessions you can observe. Flag anything that looks broader than my current jobs need.

2) SENSITIVE INFO: Search /workspace and obvious config locations for API keys, .env files, private keys, or credential exports. Quote file paths only; never echo secret values.

3) AGENT / BOT ACCESS: List active Bots, shared templates, and routines. Note any with automatic send, purchase, publish, or local-computer execution enabled.

4) APPROVAL RULES: Summarize Ask-first vs allow-automatically rules for sending, spending, publishing, routine changes, and execution on my local computer (if configured).

5) ACTIVITY TRACKING: From run history or recent computer actions you can see, note unusual patterns (many failed logins, bulk deletes, new routines I did not expect).

6) UNUSED ACCESS: Connectors, browser logins, or Bots that nothing active references; recommend revoke/sign-out candidates for my manual review.

End with a prioritized table: Finding | Risk (low/med/high) | Recommended human action | Section.

Do not implement fixes until I say go.

Which chat tools work with this prompt?

Paste the prompt into any of these assistants (free or paid tiers).

• Grok Bot

What tips make it smoother?

• Run after importing any marketplace Bot or third-party template. Treat it as sharing your whole login set.
• Pair with Ask-first rules for send, spend, publish, and routine changes before widening access.
• Use a separate Cursor user for high-risk workloads that need credential isolation (per xAI security FAQ).

Where do the claims come from?

Summaries use public sources linked below, not independent verification. How setups are picked.

• Read @Michael_Fenech_’s write-up (x.com)

  @Michael_Fenech_: Public thread framing a monthly shared-computer security audit before giving Bots more jobs.

• Read original on docs.x.ai

  Shared computer, local execution settings, connector hygiene.

• Read original on docs.x.ai

  One computer per user; shared browser, files, and credentials.

• Read @Ken_LinkWell’s write-up (x.com)

  @Ken_LinkWell: Community scan finding shared credential files; separate Bots not a boundary.