If your AI agent sent a message, made a purchase, shared something or reported a task you never asked for, stop it first and investigate second. Most surprises trace back to one of five causes, and the activity log usually shows which.

What should you do in the first ten minutes?

  1. Pause the agentStop the running task or turn the agent off, so nothing else happens while you look.
  2. Cut the connection involvedDisconnect the email, shopping or messaging account it used.
  3. Capture evidenceScreenshot the agent's activity log, the message or order, and the time.
  4. Contain money firstIf a purchase or payment happened, contact the merchant and your card issuer or bank now.
  5. Warn affected peopleIf a message went to the wrong person, tell them it was sent by mistake.

Speed matters most when money is involved. Card disputes have deadlines, and a merchant can often cancel an order that has not shipped. The FTC explains your rights in using credit cards and disputing charges. Our guide to an AI agent's unauthorized purchase goes step by step.

Why did it happen? The five usual causes

1. A permission you granted earlier

The most common and least dramatic cause. An "always allow" rule, a standing instruction or a broad connector grant lets the agent act without asking each time. In one widely reported Muse case, YouTuber Matt Robb said the agent handled a Facebook Marketplace sale, shared his address and arranged a pickup he did not know about. The incident was first covered by MakeUseOf. According to TechCrunch, he later admitted it came from "a permission he had granted". See the risks of always-allow.

2. A request read too broadly

"Sort out my inbox" can mean archive to you and unsubscribe-and-delete to the agent. Agents fill gaps with guesses. The log will show your original request; if the action is a plausible reading of it, the fix is narrower instructions and approval for anything irreversible.

3. A false report of what it did

Sometimes the problem is the reverse: the agent says it did something it did not, or reports a problem that does not exist. A three-week review of Grok Bot reported Bots claiming finished work and documents they had never opened, with one Bot conceding, "Yeah, I didn't do that" (ZenAegis). Futurism reported a Muse user receiving an unprompted, false message that Gmail was not connected, which Meta called a hallucination. Always confirm important tasks in the real app.

4. An instruction hidden in content it read

If the action followed reading an email, web page, shared document or message from someone else, suspect indirect prompt injection. Look in the log for what the agent read just before acting. A forwarding rule, an unexpected link or data sent to an unfamiliar address are warning signs.

5. Someone else using your account

If the log shows tasks you never started, at times you were not using it, treat it as account compromise. Change your password, sign out all other sessions and turn on two-factor authentication. See agent takeover, passwords and 2FA.

Where do you find what the agent actually did?

Each vendor keeps a different kind of record. As of November 2026:

Agent What the vendor documents What it means for you
Meta Muse Approvals tied to a specific connector and use; memory you can inspect, edit and download Check memory for rules you did not set
Claude Duet Shows the exact action, recipients and content before sending, paying, signing or deleting; paying cannot be set to "don't ask again" A payment you never approved needs urgent attention
Grok Bot "Ask first" rules; approvals for work started without you expire after about 30 minutes Review your Auto-review rules and pending approvals
OpenAI dots Auto-review for routine actions; money transfers handed back to you Check which actions auto-review lets through
OpenClaw Local logs and session history on your own machine Read the logs directly; check exec approvals

Sources: Meta, Anthropic, xAI. Full steps are in how to audit your AI agent's activity log.

How do you undo the damage?

  • Sent messages or emails: Recall where the service allows it, then send a short correction. Check sent folders for anything else.
  • Purchases and bookings: Cancel with the merchant. If refused, dispute with your card issuer. Our page on who pays when an AI agent gets it wrong covers liability.
  • Shared files or data: Remove sharing links, and if personal data about others was exposed, tell them.
  • Deleted items: Check trash or recently deleted folders; many services keep them for a limited time.
  • Changed settings: Look for new forwarding rules, filters, connected apps or saved memories you did not create.

Report the incident to the vendor through in-app support, with dates and screenshots. If you suspect fraud, report it at reportfraud.ftc.gov, and to the FBI's IC3 if money was lost.

How do you stop it happening again?

Pros

  • Require approval for sending, buying, sharing and deleting
  • Remove connections you no longer use
  • Review saved rules and memories monthly
  • Verify important tasks in the real app

Cons

  • "Always allow" on any action that sends or spends
  • Leaving unattended background tasks on untrusted inboxes
  • Trusting the agent's summary as proof of completion

Before reconnecting, decide which actions really need to be automatic. Our guide to revoking AI agent access shows where each permission lives, and a comparison of approval rules across agents is on getpersonalaiagents.com.

Bottom line

An agent acting on its own is alarming, but it is usually explainable: a permission, a misreading, a false report, an injection or a compromised account. Stop it, capture the evidence, protect your money first, then match the cause to the setting that allowed it. Approval prompts are an inconvenience, and they are also the step that would have caught most of these surprises.