Meta launches Muse, an AI agent capable of sending emails and making payments

Muse is deployed for adults through a dedicated application, on iOS and Android, as well as on the web and directly within WhatsApp. The service is currently limited to the American market, with Meta also planning a future launch on its connected glasses.

Unlike a traditional chatbot, the agent can continue a task after the application is closed. It can, in particular, prepare and send an email, book a trip, fill out a form, organize a calendar, or assist with an online purchase. Meta states that sensitive actions, such as sending a message or making a payment, require user validation.

The system operates within a dedicated virtual machine, called Muse Secure VM, which hosts the agent and the data necessary for its operation. A second agent, Sentinel, controls internet access and must block or submit for validation certain actions deemed sensitive.

The launch comes, however, in a context of increased vigilance regarding the security of autonomous agents. Reuters reports that internal tests revealed malfunctions, including unsolicited data transfers and, in one case, access to private photos stored on iCloud. Meta states that it strengthened protections before the launch and assures that Muse complies with its security and privacy requirements.

An agent operating within applications

According to Meta, each user chooses the services to which Muse can connect and can revoke this access at any time. For messaging, for example, it is possible to allow simple reading of emails or also their sending. The agent also keeps a history of actions performed and those it plans to execute.

For purchases, Muse can use Link, Stripe’s payment wallet. The system generates a single-use card so that the main banking data is not directly exposed to the agent. Meta also plans to add Shop Pay and an integration with 1Password.

Muse is powered by Muse Spark, the AI model that Meta presents as its most advanced for so-called agentic tasks. The group explains that the assistant can remember certain preferences, proactively make suggestions, and coordinate longer projects, while allowing the user to erase stored information.

Meta emphasizes privacy

Meta states that conversations with Muse and the data stored in its virtual machine are not shared with its advertising systems. Users can also opt out of having their interactions used to train the group’s artificial intelligence models.

The Associated Press highlights that Meta strongly emphasizes these privacy guarantees at a time when the company is seeking to convince the general public to entrust more personal tasks to an autonomous agent. The challenge goes beyond simply answering questions: Muse must be able to take initiatives, handle personal data, and interact with third-party services.

Meta plans to introduce later this year a “Confidential VM” version in which the entire virtual machine, including conversations and user data, will be encrypted with a key held only by the user. The group also indicates that Muse is expected to be integrated soon into its glasses equipped with artificial intelligence features.