Meta’s September 8 Muse announcement describes a personal agent running in a dedicated Muse Secure VM. According to Meta, a separate Sentinel agent checks outbound activity, and sensitive actions such as purchases or sending email require permission. People choose connected services and can revoke access. These are the company’s documented design claims, not an independent security assessment.
The announcement distinguishes that system from Muse Confidential VM, which it says will arrive later in the year with stronger encryption controls. That future feature should not be read as already delivered by the launch article. The same post presents the Muse app and WhatsApp as conversational surfaces, with work continuing when the user leaves the app.
Editorial take: Security terminology can sound reassuring while hiding important differences in what is available now. The useful questions are concrete: which accounts are connected, what approval appears before an action, and how is access removed? Keeping promised protections separate from shipped ones makes this agent easier to evaluate fairly.