Meta Launches Muse, an AI Agent That Can Act on Your Behalf

Meta has officially launched Muse, a new personal AI agent designed to act directly on behalf of its users, rather than simply answering questions. Available since September 9 in the United States on iOS, Android, the muse.ai website, and within WhatsApp, the tool marks a new step in the race toward agentic AI that major tech companies are waging against one another.

An Agent Capable of Carrying Out Concrete Tasks

Unlike a traditional conversational assistant, Muse can send an email, book an appointment, fill out a form, or negotiate a price on behalf of the user. Meta presents the tool as the first personal agent “designed to act” rather than simply to advise. This shift toward autonomous action echoes a trend already seen with the AI-powered online purchase payment feature tested in France in recent months.

A Dedicated Virtual Machine to Secure Data

To limit the risks tied to this autonomy, Meta runs Muse on a Muse Secure VM, a dedicated virtual machine that hosts both the agent and the user's data. An end-to-end encrypted version, where even Meta would not be able to access the data, is set to follow later this year under the name Muse Confidential VM. The company states that the agent never has access to passwords or payment methods, which are kept in a separate secure space, and that the user retains control over the permissions granted: modifying access, refusing to let conversations be used to train the models, or asking the agent to forget a specific piece of information.

This architecture addresses a growing concern in the industry: the proliferation of autonomous agents widens the attack surface of the systems they operate. Recent funding rounds devoted to securing AI agents in the enterprise reflect the same concern on the business side.

Meta Accelerates as Competition Intensifies

The launch of Muse comes amid fierce competition over autonomous agents: OpenAI opened a dedicated API in public testing in early September, while several developers are considering common standards to let these agents communicate with one another. This is not Meta's first attempt: earlier in the year, the company had already promised to give every internet user a personal AI agent, while OpenAI, for its part, chose to slow the rollout of its most powerful models for safety reasons.

• Availability: United States only for now, on iOS, Android, muse.ai, and WhatsApp.
• Features: emails, appointment booking, forms, price negotiation, online purchases.
• Security: dedicated virtual machine, no direct access to passwords or payment methods.
• Coming soon: end-to-end encrypted “Muse Confidential VM” version by the end of the year.

Frequently Asked Questions

Is Muse available in France?

No, at its launch on September 9, 2026, Muse is only accessible in the United States, via iOS, Android, the muse.ai website, and WhatsApp. Meta has not announced a timeline for a European rollout.

Can Muse access my passwords or my credit card?

No. Meta states that the agent has no direct access to passwords or payment methods, which remain stored in a secure space separate from the agent itself.

Can you prevent Muse from using your conversations to train Meta's models?

Yes, users can refuse to let their exchanges with Muse be used to train the models, change the permissions granted to the agent, or ask it to forget a specific piece of information.

Sources

• Blog du Modérateur
• Fredzone
• Prodiris

This article was written with the help of artificial intelligence. Editorial policy