Meta launched Muse, a personal agent that reads mail, runs tasks and browses the web. Each user gets a dedicated cloud machine that stores their files and login tokens. The agent runs inside an isolated compartment and never sees real credentials. A separate component called Sentinel approves every outside action and network request. Real tokens are swapped in only at the network edge. Approval prompts appear in the app itself, not in the chat with the agent. The email connector strips one-time codes, password resets and login links. Purchases use single-use card numbers tied to one merchant and amount. Meta says its model resists hidden instructions planted in data it reads. Classifiers outside the agent check for such attempts and cannot be switched off.

• up to $300,000 per valid report
• up to $130,000 for prompt injection