Meta launches Muse, a personal AI agent

Meta launched Muse, a consumer AI agent designed to carry out tasks — sending emails, booking travel, filling out forms, shopping — rather than only answering questions, available first in the United States on iOS, Android, a dedicated muse.ai site and inside WhatsApp. Meta describes it as continuing to work after a user closes the app, returning only when it needs approval to proceed.

The launch’s emphasis fell on containment as much as capability. Muse runs inside what Meta calls a Secure VM, a dedicated cloud machine that also holds the credentials for any service a user connects; a separate agent called Sentinel runs on the same machine but is kept apart at the system level, and Meta says nothing Muse does reaches the internet unless Sentinel approves it, with the user asked directly when approval is needed. Meta also says Muse has no visibility into stored passwords or payment details, including ones typed directly into a browser, which are held in separate secure storage the agent can use without seeing. The system is powered by a new model, Muse Spark, which Meta describes as its most capable model to date, built specifically for this kind of real-world agentic work.

The architecture is a direct response to a problem the industry has been confronting all year: an autonomous agent with broad tool access is also a broad attack surface, most visibly illustrated weeks earlier by OpenAI’s disclosure of an autonomous-agent breach of Hugging Face. By splitting execution from network-approval authority into two separate agents on an isolated machine, Meta is proposing a specific mitigation — supervised autonomy rather than either full autonomy or none — for consumer use, a design choice other companies building personal agents will now be measured against.