A personal AI agent is as safe as the access you give it and the approvals you keep. This checklist turns the vendors' own safety documentation and government guidance into roughly 30 concrete checks, grouped by stage: before you connect anything, while you set permissions, when money is involved, and after something goes wrong. It applies to Meta Muse, OpenAI dots, Claude Duet, Grok Bot, Gemini Spark and self-hosted OpenClaw, with agent-specific notes where the vendors differ. Print it, or keep it open while you set up.

Before you sign up: five checks

Start by reading what the vendor has actually published. A vendor that explains its permission model, credential handling and limits in writing is easier to use safely than one that doesn't.

Permissions: what should the agent be able to reach?

The most useful idea here is Simon Willison's lethal trifecta: an agent that has your private data, reads untrusted content (emails, web pages, messages) and can communicate externally can be tricked into leaking that data. You can't always remove all three, but you can narrow each one.

  • Start read-only. Where a connector offers read access separately from send or edit access, take read first. See read-only vs send access.
  • Connect one service at a time. Add the next only when you have a task that needs it. This is the least-privilege principle, which the joint CISA, NSA and NCSC guidance calls "critical".
  • Skip the accounts you can't afford to have misused. Primary bank login, work email on a personal agent, health portals and password managers are poor first connections.
  • Avoid "always allow" for anything that sends, publishes, deletes or pays. Muse offers one-time, session, task, time-bounded and perpetual approvals; pick the narrowest that works. Why "always allow" is risky.
  • Turn on per-action rules where offered. Grok Bot lets you set "Ask first" rules for sending, publishing, deleting and purchasing. dots has Custom Rules with four modes.
  • For OpenClaw, turn the sandbox on. OpenClaw 2.0's sandbox is off by default, and its docs say it is "not a hostile multi-tenant security boundary". Follow our OpenClaw hardening guide.
  • Don't treat separate bots as separate boxes. xAI's FAQ says all of a user's Bots share one cloud computer: "Do not use separate Bots as a security boundary."

Passwords and sign-ins

  • Never type a password into the chat. The hosted agents are designed so the model doesn't see credentials: Muse uses surrogate tokens, Grok Bot uses Vault or a Secure Form, and Anthropic says Duet "never sees your passwords, browser cookies, sign-in tokens, or other credentials".
  • Handle 2FA codes yourself. Grok Bot hands passwords, 2FA and payments to you through a takeover flow, and Muse's email connector filters out one-time codes and reset links. See takeover, passwords and 2FA.
  • Use a password manager and unique passwords for every account the agent touches, so one compromised session doesn't unlock the rest.

Money: payments and spending

Minimum age for Muse, Duet, Gemini Spark
18+
US window to dispute a credit card billing error
60 days
Max debit loss if reported within 2 business days
$50
  • Pay with single-use cards where the agent supports them. Muse uses single-use Link by Stripe cards; Duet uses one-time Link cards limited to one merchant and amount. Our payment security explainer covers how these work.
  • Prefer a credit card over debit. Under the Fair Credit Billing Act you can dispute a billing error in writing within 60 days of the statement, per the CFPB. Debit losses under Regulation E depend on how fast you report.
  • Set a low limit or alerts on whatever card the agent can use, and turn on instant transaction notifications.
  • Never let an agent move money between accounts. Duet blocks this even with your approval, and dots always hands money transfers back to you. More in can AI agents move money?.
  • Read every purchase approval in full: merchant, item, quantity, price and delivery address.

Data and memory

  • Check the training setting. Muse training is on by default with an opt-out, per CNBC. Anthropic has no Duet-specific training statement, but its consumer policy trains only if Model Improvement is on, and that most likely applies to Duet. Our training opt-out guide covers each.
  • Assume disconnecting isn't deleting. dots and Duet both keep what they learned from a service after you disconnect it. See deleting agent memory.
  • Don't share what you wouldn't put in an email to a stranger: full card numbers, SSNs, medical details you don't need help with. Our list of what not to tell an AI agent goes further.

Ongoing habits

  • Review the activity log weekly and after any money or messaging task. Our activity log audit guide shows where to find it.
  • Treat surprising instructions as a red flag. If the agent suddenly wants to email a new address, change a setting or buy something you didn't mention, stop and check. That is how indirect prompt injection usually shows up.
  • Vet skills and connectors before installing. Koi Security found 341 malicious skills out of 2,857 on OpenClaw's ClawHub in early 2026. See ClawHavoc.
  • Keep the app updated. The Muse zero-day reported on 21 September 2026 was fixed by a Meta update.

If something goes wrong

  1. Stop the agentPause or cancel the task, then revoke payment and send access.
  2. Save evidenceExport or screenshot the activity log, approvals and confirmation emails.
  3. Call your card issuer or bankReport the charge promptly; timing affects your rights.
  4. Report itTell the vendor's support team, and file at reportfraud.ftc.gov or ic3.gov if it involved fraud.
  5. Change passwordsRotate any password or token the agent could reach.

Whether a purchase you technically approved counts as "unauthorized" is legally unsettled, so don't assume a refund. Our guides on unapproved purchases and when an agent did something you didn't ask explain the next steps in detail. For a deeper comparison of how each agent's approval rules work, see the approval rules comparison on Get Personal AI Agents.

Bottom line

You don't need every box ticked on day one. The four that matter most are: approvals on for sending, paying and deleting; no banking passwords; single-use or low-limit cards; and a weekly look at the log. Everything else narrows the room for a mistake or an attack.