Guide from the official docs

How AI agents handle your passwords and payments

Which AI agents see your passwords and cards, which fill them in from a vault, which stop before a payment, and how 1Password for Agents fits in. From vendor docs.

By AideDex desk · published · updated

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: most of the big agents are built so the AI model never sees your password. You type it into a separate form, or a vault fills it in for the agent. Most also stop and ask before a purchase, and a few refuse money transfers outright. The exceptions matter: Instinct’s privacy policy describes giving it your username and password, OpenAI says passwords you paste separately into a chat fall outside its protections, and Google says not to enter sign-in info into a Gemini task thread.

Everything below is what each vendor says in its own docs. These are vendor claims, not independent audits, and AideDex has not tested them.

Passwords: who sees them

AgentHow sign-in works, per the vendor
ChatGPT dotsWhen a sign-in is supported, the model pauses and you enter your details in a secure login form that passes them directly to the browser environment. Saved passwords go through “a dedicated encrypted credential service” without passing through the model (OpenAI)
Meta MuseCredentials for connected services are stored on your own cloud computer, outside the part where the agent runs; the agent “doesn’t see real credentials” (Meta AI research)
Claude Duet“It never directly sees your passwords or sign-in tokens” (Anthropic)
Claude in ChromeWith 1Password for Claude, “1Password fills the login directly, and your passwords and one-time codes never enter Claude’s context” (Claude Help Center)
Gemini tasksWith your permission it can use logins saved in Google Password Manager. Google says not to enter sign-in info or payment details in the task thread, and that Gemini may ask you to “take control” of the browser to type them yourself (Gemini Apps Help)
Perplexity Comet“Comet never has access to your passwords or payment data” (Perplexity Help Center)
Opera NeonCredentials you enter when signing in to a site “are not sent to any third-party servers (such as Opera’s)” (Opera). Tools you connect over its MCP Connector work inside that signed-in session; see how to get started with Opera Neon
Hark ProPasswords and cards sit in an encrypted vault, Secured by Hark. Hark says “Even we can’t see inside” and it can’t see a password as it is filled in (Hark)
LVLZPasswords and cards are encrypted at rest, and “each workspace or user has its own encryption key”; a server-side browser fills them in and “The AI model sees that a field was filled, never the value itself” (LVLZ)
WabiCredentials, cards and IDs you save in the Wallet are encrypted and “released for use only after you authenticate on your device” (Wabi)
Grok BotFor passwords, passkeys and two-factor codes you take over the Bot’s computer, and “The Bot doesn’t see your password” (xAI Docs). Since Oct 23, 2026 Bots can also save logins and cards to a Vault in Settings (xAI changelog); see how to use Grok Bot’s Vault
InstinctThe privacy policy’s own example: you may give the assistant “your username and password for third-party accounts so that the personal assistant can sign into these accounts on your behalf” (Instinct)

Note the difference between “the model can’t see it” and “the company can’t see it”. LVLZ says the keys are derived from “a master secret that never leaves the server environment” and that a server-side browser fills in passwords (LVLZ); Hark says even Hark can’t see inside its vault (Hark).

The gap: passwords you paste in yourself

OpenAI’s protections “apply to supported sign-in flows. They don’t cover passwords you share separately in a chat or document, or through a plugin” (OpenAI Help Center). Google gives similar guidance for Gemini: keep sign-in info and payment details out of the task thread, and when they’re required, take over the browser yourself and type them on the webpage (Gemini Apps Help). If an agent asks for a password in chat, use its secure sign-in or take-over option instead.

1Password for Agents

1Password announced 1Password for Agents on Oct 22, 2026. According to the company, the person sees which credential an agent needs and why, approves the request on their own device, and 1Password enters it directly in the login form rather than passing it to the agent. Agents it names include Muse, Claude Duet, ChatGPT Work and dots, and Grok Bot; for ChatGPT it says “Users approve sign-ins, and the passwords aren’t exposed to the AI model” (1Password). For Claude Duet, the release quotes Anthropic saying “soon 1Password will complete the sign-in” (1Password). Instinct and Perplexity are among the “additional partners working on 1Password for Agents” (1Password). Anthropic already documents 1Password for Claude in Chrome (Claude Help Center), and Meta had announced 1Password support for Muse at launch as coming soon (Meta).

Payments: what needs your approval

AgentBefore it pays, per the vendor
ChatGPT dotsCan buy with a card you saved on a merchant’s site; “These purchases require your approval, which may be given in advance when it specifically covers the purchase.” You must take over for money transfers or password changes (OpenAI Help Center)
Meta MuseAsks for approval before a purchase. It can check out with Link by Stripe, whose agent wallet “generates a one-time-use card so your real card details stay hidden” (Meta)
Claude DuetShows what it will do and waits before it “makes a payment”; it won’t trade securities or move money between accounts, even if you approve. A new classifier is “designed to reject requests to share” card numbers and government IDs (Anthropic)
Claude in ChromeCan’t make purchases, create accounts or execute trades, in any permission mode (Claude Help Center)
Gemini tasksCan stop and hand control back to you so you type passwords or payment details yourself (Gemini Apps Help)
Opera NeonPayment details “stay within your browser session and are never sent to Opera or any third-party servers”; at any point you can pause the agent, stop it or take control yourself (Opera)
LVLZBefore the agent makes any payment, you see the total, the items and the merchant; the check runs in code, so the model “cannot perform it until the approval is recorded”. You can turn individual rules off (LVLZ)
InstinctUnder the terms, agreements the service enters into for you “shall be binding on you as if entered into directly by you” (Instinct)

Two of these are worth reading closely. OpenAI’s approval “may be given in advance”, so a standing instruction can count as approval (OpenAI Help Center). LVLZ’s rules can be switched off by you (LVLZ). Our pay bills and shop page lists which agents can buy things at all.

Connecting Link or Shop Pay to Duet is covered in how to get started with Claude Duet. Which brokers let an agent trade, and which agents refuse, is in can an AI agent trade stocks?.

Scheduled tasks and money

Approvals only help if you are there to give them. Anthropic advises against scheduling Cowork tasks that “make purchases, or take other actions that are difficult to undo” (Claude Help Center). Google says that if a Gemini schedule runs while you are offline, “you may not be able to stop Gemini from completing an unintended action” (Gemini Apps Help).

What you can do

Each agent’s safety page has the full data access, training and retention details, and our Gmail guide covers email access.

FAQ

Can Meta Muse see my credit card?

No, per Meta (see the table above). In October, Meta’s spokesperson added that you’re able to step in and take control whenever you want, that Muse is built to request your permission every time before it buys anything, and that it applies “ethical browsing principles” to requests a person couldn’t make, such as buying every ticket to an event (The Verge).

Can ChatGPT dots buy things without asking?

Not with a card saved on a merchant’s site, unless you approved that purchase, possibly in advance. Custom Rules let you choose which supported actions a dot may take on its own, which need approval and which it should never take, but they can’t switch off core safety requirements or Auto-review, which checks certain planned actions against your instructions before they run (OpenAI Help Center).

Does 1Password work with AI agents?

It is being built in. Whenever a login form comes up, the agent proves its identity to 1Password with a unique digital key. Besides the agent makers, 1Password names Ramp, KERNEL, Browserbase and Browser Use among the partners working on it, and its Chief Product Officer says an agent should not get “permanent access to everything behind the account” (1Password).

Sources

18 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. Meta: Introducing Muse (Sep 8, 2026) about.fb.com
  4. Meta AI research: Security and safety for AI agents, our approach with Muse research.meta.ai
  5. Anthropic: Introducing Claude Duet claude.com
  6. Claude Help Center: Claude in Chrome permissions guide support.claude.com
  7. Claude Help Center: Use Claude in Chrome safely support.claude.com
  8. Claude Help Center: Use Claude Cowork safely support.claude.com
  9. Gemini Apps Help: Tasks in Gemini support.google.com
  10. Perplexity Help Center: Control what Comet Assistant can use perplexity.ai
  11. Opera Neon FAQ operaneon.com
  12. Hark: Introducing Hark Pro hark.com
  13. LVLZ: Security lvlz.ai
  14. LVLZ: Why approvals belong outside the model (Oct 16, 2026) lvlz.ai
  15. Wabi Privacy Policy wabi.ai
  16. Instinct Privacy Policy instinct.com
  17. Instinct Terms of Service instinct.com
  18. 1Password: 1Password for Agents launches (Oct 2026) 1password.com

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