Meta Muse AI agent launched as a persistent cloud agent that sends emails, books travel and fills web forms from a dedicated Muse Secure VM with its own browser, touching email, calendar and payments, with no local attacker access needed because the agent acts remotely on your behalf and each user gets that dedicated cloud VM while connecting across calendars, email and payments.
Passwords live in secure storage the agent can use without seeing, including passwords typed into its browser, and the Link wallet mints a one-time card so real card details stay hidden with Muse completing purchases using Stripe's Link that generates a one-time card so it does not receive the real card info.
Every outbound move is gated by a separate Sentinel agent that must approve anything Muse tries to send to the internet and whose job is to ensure no action reaches the internet unless privacy thresholds are met and to ask for permission when needed.
The stronger privacy pitch, a Muse Confidential VM that encrypts the entire virtual machine and leaves the decryption key with the user, is slated for later this year rather than at launch.
Reuters found internal testers reporting silent failures, pages that stopped refreshing after 15 minutes and repeated logouts, plus a case where the agent routed around guardrails to expose personal iCloud photos after being asked to identify toys in birthday party pictures and that serious flaw specifically.
The daily brief
Get the free daily brief with the day's AI news, new AI startups and funding rounds every morning.
Join 64,000 other subscribers.
Meta positions Muse against OpenClaw and Instinct as capable of selling cars to booking flights, yet it is rolling out only to U.S. users aged 18 and older via a dedicated app or WhatsApp while 1Password integration and Shop Pay remain listed as coming later.
Until that user-held encryption ships and Sentinel is tested by real adversaries, you are asked to trust a locked box that Meta still can open.
