If an AI agent asks you to paste a password or a two-factor code into a chat, stop. Legitimate personal agents are built so the model never sees your secrets. Instead, they pause and hand control back to you through a takeover, a secure sign-in form or a password manager. A request for credentials in chat, by text or by email is far more likely to be a scam than a real agent. This guide explains how each agent's handover works, based on official docs as of November 2026.

Why do agents hand control back at all?

Three kinds of step are deliberately kept with a human:

  • Credentials. Passwords and sign-in tokens are the keys to your accounts. If the model saw them, a prompt injection could try to make it leak them.
  • Second factors. 2FA codes and magic links prove that you, not someone with your password, are signing in. An agent that could read and use them would defeat the purpose.
  • Human checks and payments. CAPTCHAs confirm a person is present; payments need your consent.

So when your agent stops and asks you to take over, it is working as designed.

How does each agent handle passwords and 2FA?

Agent Passwords 2FA, CAPTCHAs, payments
Meta Muse Muse sees only surrogate tokens, not passwords or payment methods Email connector filters one-time codes, reset links and magic links
OpenAI dots Saved passwords go through a secure form, not exposed to the model Password changes go back to you; "Take over" on the cloud computer
Claude Duet Never sees passwords, cookies or sign-in tokens; a proxy swaps tokens Sign-in form appears only in the app
Grok Bot Credentials via Vault, a password manager or a Secure Form Passwords, 2FA, CAPTCHAs and payments go through takeover

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

Meta Muse

Meta's security post says Muse only ever handles surrogate tokens, so it cannot see your passwords or payment methods. Its email connector also filters out one-time codes, password-reset links and magic login links, so an injected instruction to "forward the latest code" has nothing to forward.

OpenAI dots

OpenAI says saved passwords are entered through a secure form and aren't exposed to the model, and that password changes always go back to you. You can watch a dot's cloud computer in the Activity View and press "Take over" to step in, per VentureBeat's launch report. eesel's review noted a dot would not fetch a 2FA code from an inbox it could read. That is annoying in the moment but is the protection doing its job.

Claude Duet

Anthropic says Duet never sees passwords, cookies or sign-in tokens; a proxy swaps tokens in. It is also specific about where sign-in happens: the form appears only in the app and "never arrives as a link, a text message, an email, or a website other than claude.ai", according to its safety article. You cannot yet watch or take over Duet's browser directly.

Grok Bot

xAI's FAQ says passwords, 2FA, CAPTCHAs and payments go through a takeover flow, and credentials are passed via Vault, a password manager or a Secure Form. Remember that all your Bots share one cloud computer, including logins, so a session you sign into for one Bot may be usable by others. See Grok Bot computer access settings.

How do scammers exploit this?

Because agents really do ask you to step in, scammers imitate the request. Common patterns:

  • A text or email saying "Your agent needs you to verify your account" with a link.
  • A fake sign-in page styled like the agent's app. See fake AI agent login pages.
  • A message in a chat app claiming to be your agent, asking for a code "to finish the booking".
  • A web page the agent visits that shows text addressed to you, asking you to paste a code.

The tell is the channel. Real handovers happen inside the agent's own app or desktop client. Claude Duet states this outright. For texting scams in general, our sister site covers AI text scams.

What is the safe way to complete a takeover?

  1. Open the app yourselfGo to the agent's app or desktop client directly, not via a link in a message.
  2. Check the requestConfirm the site and task match what you asked for.
  3. Use the secure screenEnter credentials only in the vendor's form, vault or password manager.
  4. Finish the human stepSolve the CAPTCHA or approve the 2FA prompt on your own device.
  5. Hand back and watchLet the agent resume, then check its activity log.

What if I already shared a password or code?

Act quickly and calmly:

  1. Change the password from the service's own website or app, not from a link.
  2. Reset or re-enrol two-factor authentication and sign out other sessions.
  3. Check the account's recent activity, sent mail and forwarding rules.
  4. If a bank, card or payment account is involved, call the number on the back of your card.
  5. In the US, report fraud at reportfraud.ftc.gov or, for internet crime, to IC3. Elsewhere, use your national fraud reporting service.

Then remove the agent's access to the affected account using our guide to revoking AI agent access, and reconnect only through the official flow.

Bottom line

Handovers are a safety feature, not a bug to engineer around. Do not look for ways to let an agent read codes or type passwords for you. Keep credentials in the vendor's secure flow or a password manager, and treat any request for secrets in chat as suspicious. For more on what agents can and cannot see, read does an AI agent see my passwords, and for which tasks to keep for yourself, see what to let an AI agent do.