Every major personal AI agent keeps some record of what it did. Meta Muse has an audit trail of past and planned actions, OpenAI dots has the Activity View in its desktop app, Claude Duet keeps approvals in the main thread and the Tasks tab, Grok Bot has a usage dashboard, and OpenClaw logs whatever you configure. A ten-minute weekly review of sent messages, purchases, new connections and standing approvals catches most problems while they are still small. This guide shows where each log is and what to check, based on official docs as of November 2026.

Why audit an agent at all?

Approval prompts only catch what you are asked about. Some actions run without asking because you set a rule, gave a standing approval or chose a mode that allows it. Others happen in tasks started on a schedule. The log is the only place you see those.

Auditing also catches slow problems: a connector you forgot about, a perpetual approval from months ago, or a pattern of small purchases. Meta's help centre says Muse "may be inaccurate or take unexpected actions", and every vendor says prompt injection is reduced, not eliminated. The log is how you notice.

Where is each agent's log?

Agent Where to look Live view? Notes
Meta Muse Audit trail of past and planned actions Not documented Shows planned actions too
OpenAI dots Activity View in the desktop app Yes, plus "Take over" Monitor, redirect or stop
Claude Duet Main thread and Tasks tab No browser view yet Unanswered approvals come back
Grok Bot grok.com/dashboard → Usage Takeover for sensitive steps Enterprise plans add audit controls
OpenClaw Gateway logs and session history you keep Control UI Run openclaw security audit

Based on vendor documentation and launch coverage, as of November 2026.

Meta Muse

Meta's security post says Muse keeps an audit trail of past and planned actions. The planned part is useful: you can see what Muse intends to do before it does it. Meta does not document the exact menu path in the sources we reviewed. Check approvals here too, since standing approvals are what let actions skip you. See Meta Muse safety settings.

OpenAI dots

The Activity View in the desktop app lets you monitor what a dot is doing, redirect it or stop it, and you can watch its cloud computer and press "Take over", per VentureBeat. On a work account, workspace owners control capabilities, which may include what is visible.

Claude Duet

Approvals appear in the main thread and the Tasks tab, and return if you do not answer, according to Anthropic's help centre. Only you can mark a task complete, so review open tasks for anything Duet is still working on. You cannot yet watch Duet's browser live.

Grok Bot

Usage is at grok.com/dashboard → Usage. Also check the shared cloud computer for files and logged-in sessions, since all your Bots use the same one, per xAI's FAQ. Check the drafts folder for each Bot's @mail.grokbot.com address, where unrequested outbound mail waits.

OpenClaw

OpenClaw has no vendor-side log. Keep gateway and session logs, review exec approval history in the Control UI, and run openclaw security audit to spot configuration drift from conservative defaults, per its security docs.

What should I check each week?

  1. Sent itemsScan messages and emails sent by the agent for unknown recipients or private details.
  2. PurchasesMatch every charge to a task you asked for; check single-use card activity.
  3. New connectionsLook for connectors, apps or logins you do not remember adding.
  4. Standing approvalsRemove perpetual or always-allow permissions you no longer need.
  5. Unattended tasksReview anything that ran on a schedule or while you were away.

Also check the other side

The agent's log shows what it thinks it did. Cross-check with the accounts themselves:

  • Your email's Sent folder and forwarding rules.
  • Your card or bank statement.
  • Your Google or Microsoft account's list of apps with access.
  • Calendar invites sent from your account.

A forwarding rule you did not create is a serious red flag, whatever the agent's log says.

What counts as a red flag?

Pros

  • Actions that match tasks you started
  • Approvals you remember giving
  • Purchases matching single-use card charges

Cons

  • Messages to people or addresses you do not recognise
  • New connectors, filters or forwarding rules
  • Tasks that ran at odd hours you did not schedule
  • Deletions of email, files or calendar events

If something looks wrong, see what to do if your AI agent did something you didn't ask.

What should I do if I find something wrong?

  1. Save the evidence. Screenshot or export the relevant log entries with times before changing anything.
  2. Stop the agent. Pause or stop the task (dots' Activity View has a stop control) and revoke access using our guide to revoking AI agent access.
  3. Fix the account. Change passwords from the service's own site, remove forwarding rules and sign out other sessions.
  4. Contact the money side. For an unexpected charge, contact your card issuer promptly. The CFPB says credit card billing disputes should be sent within 60 days of the statement. Outcomes are not guaranteed, especially for purchases you approved. See unauthorized AI agent purchases.
  5. Report. Tell the vendor's support. In the US, report fraud at reportfraud.ftc.gov or internet crime to IC3.

Bottom line

Approvals cover the moments you are asked about; the activity log covers everything else. Check it weekly, cross-check with your accounts, and remove standing approvals you no longer need (see why "Always allow" is risky). For real-world cases of what can go wrong, see our sister site's personal AI agent incidents tracker.