Meta launched Muse on 8 September 2026, a personal AI agent rather than a chat. It sends emails, books travel, fills in forms and checks out with single-use payment cards. There is a free tier, and paid plans at 20 and 100 dollars a month. Meta says more than 10 million people were using it by its third-quarter results.
It learns from you unless you say no
WIRED reported that users are opted in to having Muse improve Meta's AI. To stop it: Settings, Data controls, turn off "Help improve our AI models". There is no switch to turn Muse's memory off; you can ask it to wipe or edit what it remembers.
It asks for a lot
In ZDNET's hands-on test, Muse asked for full access to the computer's disk, the microphone and contacts, and offered to store passwords. WIRED said the app kept nudging users to connect bank, email and passport details. ZDNET later noted that people who never signed up are exposed too: details of your contacts and meeting guests can flow into a user's requests.
The first two months
Within weeks a security researcher showed a way to hijack the agent, and Meta's own documents call this kind of attack an open problem. A columnist said Muse synced his Mac messages with disk access off; Meta denied it. In October Meta paused Muse's sharing feature after researchers found shared-chat links were predictable and visible in public records, exposing about 6,800 conversations.
If you try it
Turn off the training switch first. Connect only the accounts a task needs, and disconnect them after. Do not let it store passwords. Keep payments to the single-use cards and read every action before you confirm. And do not share chats.