1Password on Thursday introduced a platform that lets personal AI agents from Meta, OpenAI and Anthropic sign into a user's accounts once the user approves the request. The agent then carries out the task without ever seeing or storing the password.

The product, called 1Password for Agents, inserts a verification step between the agent and the login page. When an agent taps into a sign-in form, it presents a unique digital key that 1Password uses to authenticate its identity. The user is shown which account is being requested and why, and approves on their own device. 1Password then fills the credential directly into the form, the agent proceeds with the task, and the password remains in the vault.

"Agents are starting to do real work on our behalf," said Aashish Tripathi, chief product officer at 1Password. He added that a single task assignment should not give an agent permanent access to everything behind an account.

Several major AI companies are already integrating the platform. Meta's Muse will route logins through 1Password once the integration is enabled, while OpenAI has wired up both ChatGPT Work and its always-on Dots agents. Elon Musk's SpaceXSI signed up Grok Bot, and Town, an AI assistant startup that raised $108 million this week, has also joined.

Anthropic's Claude has used 1Password in the browser since July, and the company's more recent Claude Duet agent, released on Wednesday, comes next. "Claude Duet asks before it signs in anywhere, and soon 1Password will complete the sign-in with the passwords people have already set up, so they don't have to start over," said Ami Vora, chief product officer at Anthropic. Vora said Anthropic wants everyone to be able to use the feature, even those "who've never before let an agent near their accounts."

Nine more companies, including Perplexity and Browserbase, are building their own integrations. Most of that adoption has yet to go live: Meta's Muse integration still needs to be enabled, and Claude Duet's is still to come.

US banks detect agent activity inside customer accounts

The launch came on the same day that reports emerged of US banks observing agent activity inside customer accounts, with most banks unable to tell whether that activity traces back to a genuine user request. Chase Bank has told customers not to give their usernames and passwords to anyone. 1Password noted that typing a password into an AI chat saves a copy the user cannot control, and said its new system is designed to prevent exactly that.

Security concerns around AI agents have been building. During a June assessment, an OpenAI agent accessed an Australian Medicare statistics portal without authorization, as Cryptopolitan reported. Prime Minister Anthony Albanese has directly complained about a delay of about three months in the disclosure.

Meta's Muse 10.2 million US downloads, but researchers atoc also identified more than 6,800 shared-chat links that anyone could access. The incidents point to exposure from both directions: an agent acting without authorization, and shared-chat links that put conversations within anyone's reach.

"As agents take on more complex tasks for users, security and user control need to be built in from the start," said Glen Coates, head of app platform at OpenAI.