AI agents have a trust problem: to log into a site they need the password, and once they have it, it can end up anywhere. 1Password is trying to remove that step with 1Password for Agents, a feature that lets agents access your accounts without ever seeing or handling the credential.
Here is how it works. When the agent reaches a sign-in form, it uses a unique digital key to prove its identity to 1Password. The app then sends you a notification saying which account is needed and why. If you approve, 1Password fills in the login fields directly, without handing anything to the agent, which carries on with its task. Approval happens on your own device, similar to two-factor authentication, so access is not granted automatically across every connected account.
The system is designed to work with several widely used agents, including Meta Muse, ChatGPT Work and Dots, Claude Duet, Grok Bot and Town. 1Password says other partners are working on integration, including Browserbase, Browser Use, Dana, Instinct, Kernel, Perplexity, Ramp, Tomo and Wajo, with "additional integrations to come."
The move responds to a problem that has surfaced in recent weeks. With ChatGPT Work, credentials are stored in cookies in the cloud browser so logins can be handled autonomously, a convenience that carries the risk of hijacked active sessions. 1Password's model instead aims never to leave the credential in the agent's context, a logic already discussed in an article on how to stop AI agents from exposing API keys.
Image: dteare — https://github.com/dteare/raycast-1password-extension
How well the promise holds in practice remains to be seen. Control shifts to user approval for each request, so security also depends on how carefully users actually read what they approve.
No comments yet. Be the first!