Meta unveils Muse, a public AI agent with monitored VM

Meta announces Muse, a personal agent designed for everyday use and immediate accessibility. The company promises execution within a virtual environment monitored by another agent, Sentinel, along with data control options. The U.S. launch is set to be free for most users, with subscriptions to follow, the details of which remain unclear.

Isolated Execution, Sentinel Monitoring, and Regulated Payments

Muse operates in a cloud-hosted virtual environment designed to separate and secure each user's data. Meta specifies that Muse does not have access to passwords or payment methods, and a distinct AI agent, Sentinel, monitors the virtual machine to prevent any actions by Muse from reaching the Internet without approval. A so-called "confidential" version of this virtual machine, encrypted to remain inaccessible even to Meta, is planned for later this year. Support for 1Password and Shop Pay is in the works, while Muse can already make secure payments via Stripe’s Link. Meta emphasizes the ability for users to opt-out of having their interactions used to train the company’s AI models, without clarifying whether this option is enabled by default, and allows users to request that Muse forget specific information. These measures aim to address security and privacy risks related to potential access to sensitive data.

A Credibility Challenge Amid Past Incidents

Meta needs to regain trust after recently facing a lawsuit regarding the dangers of social media and being involved in privacy scandals, including the Cambridge Analytica incident. The company's AI initiatives have already encountered issues, such as a Discover feature that exposed content from other users, a support chatbot that allowed hackers to take control of over 20,000 Instagram accounts, and an AI-generated news feed that appeared to violate its own rules. Additionally, there is a recent reputation tied to "creepy glasses." These past incidents pose a hurdle for establishing Muse as a trusted assistant.

Promised Features: Autonomy, Contextual Memory, and Messaging

Meta presents Muse as an agent capable of executing end-to-end tasks, such as opening a browser, filling out forms, or negotiating on behalf of the user. For longer tasks, Muse continues to work in the background after the application is closed and prompts the user in case of changes or required approvals. The agent remembers details shared only once to offer suggestions later without prompting. Interaction occurs via the Muse app or WhatsApp. Meta highlights usability without technical experience or a learning curve, for everyday tasks such as online shopping, sending emails, or planning trips.

U.S. Launch, Majority Free Access, and Undetailed Subscriptions

The launch of Muse is planned in the United States on iOS, Android, and the website muse.ai, with AI glasses announced as "coming soon." The assistant will be free for most users, but Meta plans paid subscriptions for those seeking additional features. The company has not specified the nature or cost of these paid offerings, nor defined the limits applicable to free accounts.

Strategy, Internal Model, and Competitive Landscape

Muse is presented as the core of Meta's AI strategy for the future, based on the internal model Muse Spark, as part of a multi-billion dollar strategic overhaul. Meta describes Muse as a personal AI agent designed for everyone, but the product enters a market already occupied by competing assistants and agents, such as ChatGPT Work, Claude Cowork, Copilot Tasks, Grok Bot, and Moltbot, often aimed at businesses or advanced users. Google is also targeting a similar audience with Gemini Spark. Meta is banking on its brand recognition and position as a social media giant to reach a broad audience, presenting Muse as a first step toward an agent that can handle more tasks, allowing users to focus on what matters to them.