1Password launched 1Password for Agents on Oct 22. When a supported personal agent reaches a sign-in form, it identifies itself to 1Password with its own key and asks for a credential. The user approves or denies the request on their own device, and 1Password fills the password or one-time code directly into the page. The agent gets only metadata and a success or failure status, never the secret or vault-wide access.

• 1Password's security notes for the Claude integration say every field of an agent's request is treated as an untrusted claim, so nothing in a request (a prompt-injected reason, for example) can trigger a release by itself. There are no standing approvals, and released items are scoped to one session and discarded after it ends or a hard time limit.
• Launch partners named by 1Password include Meta's Muse, Anthropic's Claude, OpenAI, Grok Bot and Town. It is free for Individual and Family accounts on 1Password.com; .eu and .ca accounts aren't supported yet.

Why it matters: pasting passwords into an agent's prompt is one of the most common ways credentials end up in model context, logs and transcripts. Per-request human approval with fill-in-page is the pattern to look for, but approval fatigue is the obvious weak spot, so check what each request shows before approving it.