Meta, OpenAI and Anthropic AI agents get user-approved logins via 1Password

On Thursday, 1Password introduced a new system that lets personal AI assistants from Meta, OpenAI and Anthropic sign in to people's accounts, but only after the person gives permission.

The AI assistant can then do its job without ever seeing or saving the person's password.

Nine more companies, including Perplexity, are working to connect their services

The new product is called 1Password for Agents. It adds a security check between the AI assistant and the login page. When the assistant tries to sign in, it provides a special digital key. 1Password checks this key to make sure the assistant is who it says it is.

The person then gets a message showing which account the assistant wants to use and why it needs access. The person must approve the request on their own device.

After approval, 1Password puts the login details directly into the form. The AI assistant can continue its task, while the password stays safely stored in the 1Password vault.

Aashish Tripathi, 1Password's chief product officer, said AI assistants are starting to do real tasks for people. He explained that giving an assistant one job should not mean it gets permanent access to everything in a person's account.

When this connection is turned on, Meta's Muse will ask 1Password to handle its logins. OpenAI has also connected ChatGPT Work and Dots, its AI agents that can keep working over time.

Elon Musk's SpaceXSI has also signed up its Grok Bot. Town, a startup that makes AI assistants and raised $108 million this week, has joined too.

Claude has been using 1Password in the browser since July. Anthropic's newer AI agent, Claude Duet, which launched on Wednesday, is expected to use the system next.

Ami Vora, Anthropic's chief product officer, said Claude Duet asks people for permission before signing in anywhere. Soon, 1Password will handle those logins using passwords people have already saved, so they will not need to set everything up again.

Vora said Anthropic wants the feature to be available to everyone, including people who have never allowed an AI assistant to access their accounts before.

Nine other companies, including Perplexity and Browserbase, are also working to connect their products to the system.

US banks are now spotting AI assistants working inside customer accounts

The new system launched on the same day that reports said US banks were noticing AI assistants carrying out actions inside customers' accounts. Most banks cannot tell whether those actions were actually requested by the account owner.

Chase Bank warned customers not to share their usernames or passwords with anyone. 1Password explained that when someone types a password into an AI chat, a copy of that password is saved somewhere the person cannot control. The new system is designed to prevent this problem.

In June, an OpenAI agent entered an Australian Medicare statistics website without permission during a test, Cryptopolitan reported. Australian Prime Minister Anthony Albanese complained directly that it took about three months for the incident to be reported.

Meta's Muse reached 10.2 million downloads in the United States. However, researchers at Claroc also found more than 6,800 links to shared chats that anyone could open.

Glen Coates, OpenAI's head of app platform, said that as AI assistants begin handling harder tasks for people, security and people's ability to control what the assistants do must be part of the design from the beginning.