Guide from the official docs

Can an AI agent log into your bank? 2FA codes and what vendors say

Which AI agents hand the 2FA code or passkey step back to you, which steer you away from bank sites, and why the FTC says never share a code.

By AideDex desk · published

This guide follows the official docs and announcements, linked inline. AideDex did not test these products and does not score them. How we check facts.

Last checked Oct 24, 2026.

Short answer: technically some agents can sign in to a bank website, but almost none of them handle the two-factor step themselves. ChatGPT and Grok Bot pause and hand control to you for passwords, one-time codes and similar checks, and Muse lets you take over its browser at any time. Anthropic advises against using Claude in Chrome on financial accounts at all. And the bank may refuse anyway: JPMorgan Chase says it blocks outside agents from logging in with customers’ credentials.

This guide covers the 2FA and hand-back moments. For how agents store passwords and approve purchases, see AI agents, passwords and payments. Everything below is what each vendor, bank or regulator says. AideDex has not tested these flows.

Who types the one-time code

AgentWhat happens at a 2FA or security check, per the vendor
ChatGPT dots“You may need to provide an authentication code or take over to complete a security check.” Changing a password or transferring money requires you to take over and do it yourself (OpenAI Help Center)
ChatGPT Work (cloud browser)Pauses at a supported sign-in page; you enter your username and password in a secure sign-in form and “may also need to enter a two-factor authentication or security code” (OpenAI Help Center)
Grok Bot“For passwords, passkeys, two-factor codes, CAPTCHAs, and payment confirmations, the Bot should hand you control of the computer” (xAI Docs)
Meta MuseYou can take over its browser at any time; while you control it, “the agent is paused and can’t act at all” (Meta AI research)
Claude in ChromeAsks permission before accessing financial sites. With 1Password for Claude, 1Password fills logins and one-time codes so they “never enter Claude’s context” (Claude Help Center)
Perplexity CometRuns in your own browser; you can block your bank’s site so Comet won’t use it for tasks (Perplexity Help Center)
Opera NeonOpera says it blocks its agents from high-risk pages such as banking sites, and that Neon generally pauses for you to confirm a transaction (Opera); setup is in how to get started with Opera Neon

ChatGPT: secure sign-in and take over

OpenAI says that for supported sign-ins, the dots model “is paused while you complete a secure login form”, which sends credentials straight to the browser environment (OpenAI). Some steps go further than a code. Changing a password or moving money between financial accounts must be handed back to you, and Custom Rules can’t turn that off (OpenAI Help Center).

In ChatGPT Work’s cloud browser, an extra review model checks each sign-in request and where your details will go “for signs of phishing or deception”. You can see the website address and a preview of the form first. OpenAI says the username and password entered there are not visible to the model (OpenAI Help Center). Its safety list is blunt: “Never paste passwords, security codes, or payment details into the conversation.” If ChatGPT gets stuck, it asks you to take over the cloud browser on your phone or computer (OpenAI Help Center). Setting up Work itself is covered in how to get started with ChatGPT Work.

The older ChatGPT agent, with its “Take over browser” option, is “no longer available”; OpenAI points users to ChatGPT Work and the cloud browser instead (OpenAI Help Center).

Grok Bot: take over, then “I’m done”

xAI gives exact steps for the sensitive step: choose Take over on the request (or open the computer and choose Take control), complete the step, then choose I’m done and tell the Bot to continue. It adds: “Do not send a password or one-time code in ordinary chat” (xAI Docs).

Two details matter for bank logins. If you use a hardware security key such as a YubiKey, the Bot’s browser can use a key plugged into your desktop while Use hardware security keys is on; xAI says every use asks you to approve it first, and it isn’t supported on Linux yet (xAI Docs). And a site “may still block automation, require a new login, present a CAPTCHA, or require human confirmation”; the Bot “should hand those steps to you rather than bypassing them” (xAI Docs).

On Oct 23, 2026 xAI added a Vault in Settings where Bots can save your logins, cards and contact info and fill them in to sign in or check out (xAI changelog). The changelog doesn’t say whether the Vault stores one-time codes. How to add, edit and remove items is in how to use Grok Bot’s Vault.

Muse: codes filtered out of your inbox

Meta makes a specific point about email: your inbox gets one-time codes and password reset links, so an agent that reads it could pose as you elsewhere. Muse’s email connector “filters out one-time tokens, password reset links, and login magic links” with fixed filters and a classifier model (Meta AI research). Usernames and passwords you enter in Muse’s sign-in screen go to a separate credential store and are “not visible to your main agent” (Meta AI research). Meta doesn’t say how Muse handles a text-message code sent to your phone. Our Gmail guide covers inbox access more broadly.

Claude: keep banking out

Anthropic “strongly” advises against using Claude in Chrome for “Managing financial accounts or investments” and recommends a separate browser profile without access to banking, healthcare or government accounts (Claude Help Center). Claude takes screenshots of the tabs it works in and “can’t filter sensitive content out of what it sees”. It is prohibited from stock trading, bypassing captchas and “Inputting sensitive data” (Claude Help Center).

Some brokers now offer agents a separate, dedicated route instead; see can an AI agent trade stocks?.

The 1Password integration is the exception for logins. 1Password supports logins and one-time codes (TOTP); you approve each request with biometrics, and access ends when the task is done. It is in beta, on Claude Desktop for macOS only (Claude Help Center).

For Cowork, Anthropic says to block sensitive apps such as “healthcare portals, banking, dating apps” when Claude uses your computer, and to be deliberate about sites “that handle money or personal information” (Claude Help Center).

After the code: sessions stay signed in

Entering the code once may not be the end of it. In ChatGPT’s cloud browser, authentication “will persist for future tasks until it expires”; you can clear it in Settings > Cloud browser under Browser data (OpenAI Help Center). Every Grok Bot on your account shares one cloud computer, including its browser sessions and logins, and xAI warns: “Do not use separate Bots as a security boundary.” Deleting a Bot doesn’t remove those sessions; xAI says to sign out of a service when it should no longer be available (xAI Docs).

What banks say

JPMorgan Chase blocks external agents from operating on behalf of retail customers. A spokeswoman told CNBC: “Because it is not safe, we do not permit third parties to access accounts by obtaining customers’ banking credentials and logging in on their behalf.” Read-only data such as balances can still reach agents through Plaid or another connector, but agents can’t send money or pay a bill (CNBC). Jamie Dimon added that people using outside agents “should understand there’s a liability transfer” (CNBC).

Six banks, including Bank of America and Capital One, published principles for agentic commerce in September, noting consumers fear agents may “lose their money to scams and fraud” (ABA Banking Journal). Other banks and sites with agent blocks: which websites block AI agents.

What the FTC and Take Five say about sharing codes

Neither page below mentions AI agents, but both are clear about codes. The US Federal Trade Commission says: “Never give your verification code to someone else. It’s only for you to log into your account. Anyone who asks you for your account verification code is a scammer” (FTC). In the UK, the Take Five to Stop Fraud campaign says: “Treat your one-time passcode as carefully as you would your PIN” (Take Five).

That lines up with the vendors. OpenAI and xAI both tell you to keep codes out of the chat (OpenAI Help Center, xAI Docs). Anthropic says that if Claude starts “requesting sensitive information”, stop the task, as it could be a prompt injection (Claude Help Center). More on that attack: can a web page or email hijack your AI agent?

What you can do

Which agents can pay at all: shop and pay.

FAQ

Can ChatGPT enter my bank’s 2FA code for me?

Not on its own. A dot’s stored context keeps no credentials, images or screenshots, and Custom Rules can’t switch off the rule that you take over for the most sensitive actions, such as changing a password (OpenAI Help Center).

Does Grok Bot work with passkeys?

xAI lists passkeys among the steps where the Bot should hand you control of the computer. Hardware security keys plugged into your desktop work on macOS and Windows, with approval each time (xAI Docs).

Is it safe to give an AI agent my one-time passcode?

Keep it out of the chat. OpenAI says its secure sign-in protection covers supported sign-in forms only, not passwords you share separately in a chat, a document or a plugin (OpenAI Help Center). OpenAI and xAI both say not to put codes in the chat (OpenAI Help Center, xAI Docs).

Sources

17 pages
  1. OpenAI: How we build safety, security and privacy into dots openai.com
  2. OpenAI Help Center: Dots privacy, security and safety FAQs help.openai.com
  3. OpenAI Help Center: Using cloud browser in ChatGPT help.openai.com
  4. OpenAI Help Center: ChatGPT agent help.openai.com
  5. Claude Help Center: Use Claude in Chrome safely support.claude.com
  6. Claude Help Center: Get started with 1Password for Claude support.claude.com
  7. Claude Help Center: Use Claude Cowork safely support.claude.com
  8. Perplexity Help Center: Control what Comet Assistant can use perplexity.ai
  9. Opera Neon: Understanding agentic browser security operaneon.com
  10. xAI Docs: Grok Bot approvals, security, and privacy docs.x.ai
  11. xAI Docs: Grok Bot FAQ docs.x.ai
  12. xAI: Grok Bot changelog (0.75.5, Oct 23, 2026) x.ai
  13. Meta AI research: How we built safety into Muse research.meta.ai
  14. CNBC: JPMorgan Chase blocks outside AI agents like Muse, Instinct from directly accessing accounts (Oct 13, 2026) cnbc.com
  15. ABA Banking Journal: Banks release principles for building public trust in AI agents (Sep 24, 2026) bankingjournal.aba.com
  16. FTC Consumer Advice: What's a verification code and why would someone ask me for it? consumer.ftc.gov
  17. Take Five to Stop Fraud: Card fraud takefive-stopfraud.org.uk

Spotted something wrong or out of date? Report an error · Corrections log