1Password for Agents: AI can sign in without your password

Personal AI agents are getting good at booking, buying and filling in forms, but they keep hitting the same wall: the login page. On Thursday 1Password announced 1Password for Agents, a platform meant to let agents get past that wall without you pasting a password into a chat. The details below come from 1Password's announcement and 9to5Google's report.

The problem it tries to solve

Until now, when an agent reached a sign-in form you had two poor options, as 1Password puts it: take over the session and type the credentials yourself, or hand the password to the agent in a prompt, giving it access that may outlast the task. The first is slow; the second means a secret now lives in an AI conversation.

How the sign-in works

According to 1Password, the flow looks like this:

1. The agent reaches a login form and identifies itself to 1Password with a unique digital key.
2. You see which credential it wants and why, and decide whether to approve.
3. If you approve, 1Password fills the login directly into the website, not into the agent.
4. Once the site accepts the sign-in, 1Password removes the credential, so the agent can't read or reuse it as it carries on.

The company says the rest of your vault stays inaccessible to the agent, and that passwords and one-time codes aren't exposed to the AI model.

Which agents support it

1Password lists integrations rolling out "over the coming days" for Muse from Meta and Grok Bot, with Anthropic's Claude Duet, OpenAI's ChatGPT Work and dots and Town's Townie marked as coming soon. A second batch, including Perplexity, Browserbase, Browser Use and Ramp, is also listed as coming soon. Developers building their own agents can apply to integrate.

What it doesn't change

This narrows what an agent can see, but you are still trusting the agent to act sensibly once it's signed in: it can still browse, buy or send things inside that account. Approving each request is the real safeguard, so read the prompt before tapping yes. Our ChatGPT privacy settings guide covers what assistants keep about you, and Anthropic's Claude Duet preview shows the kind of background agent this is built for. For a reminder of why credentials should never travel in plain text, see the Tor exit node attack reported earlier today.