Meta's Muse agent runs tasks in a Secure VM with Sentinel checks

Muse arrives as chat AI shifts toward agents that can carry out multi-step work. Meta positions Muse around long-term personal goals — email and calendar management, travel and shopping, and cross-app chores like turning saved Instagram recipes into shopping lists — rather than one-off answers. To do that, Meta gives each user an independent Muse Secure VM, a cloud computer for browser and tool operations, and separates Sentinel at the system level to check outbound actions. This design addresses prompt injection, where a webpage or email contains instructions that conflict with the user's own; Meta says Muse does not directly see actual credentials and requires approval before actions like sending email or making purchases. Safeguards do not guarantee zero errors or leaks, and actual coverage depends on connected services, plan, and region. Compared with ChatGPT Work's cloud browser and Grok Bot's role-based bots sharing a cloud PC, the difference is less about whether an AI can operate a browser and more about what work is delegated, in which environment, and under what cost and permission conditions. On availability, Meta's Japanese newsroom says Muse was announced for the U.S. in September 2026 with Japan timing undecided, while the product page notes a free tier with usage limits and a paid plan option.