Meta Launches Muse AI Despite Internal Safety Concerns

Meta introduced its Muse AI agent in the United States on Tuesday, September 8, enabling users to have emails sent, a car sold, and travel booked on their behalf without having to be present. The agent, which is known internally as Hatch, plays a key role in Mark Zuckerberg’s initiative to deliver what he refers to as “personal superintelligence” to Meta’s user base. Each instance of Muse operates within its own cloud-based virtual machine so that it can keep working in the background after the user has moved on.

How Muse Operates and Access Options

Each Muse instance works within its own virtual machine, which is a cloud-based simulation of a personal computer, enabling it to continue operating in the background. The system has been designed by Meta using OpenClaw, an open-source AI agent framework. Muse has been developed to be able to interact with a number of app categories such as the calendar, email, payments, health, shopping, and the smart home.

You can access Muse either via a standalone app or through WhatsApp, and at present the rollout is being limited to users in the US. Meta stated that the agent will soon be available on its Ray-Ban smart glasses but did not provide a specific timeline. The basic level of the service is free, while there are two paid options, costing $20 and $100 per month respectively, for users who require more of the service. Meta has not specified the differences between the three tiers other than in the way it has described them.

Internal Testing Revealed Reliability Problems

Vishal Shah, who is Meta’s vice president of AI products, stated that the company had postponed Muse until it could deal with the security concerns before releasing it on its original April date. He added that the extra time had caused the product to meet the minimum standard that the company required in order to release it to users. Shah also admitted that the system would not be without errors, stating that the company had designed each layer of Muse’s architecture to be as safe, secure, and private as was possible.

Before the launch, internal testing revealed a number of reliability issues, such as the agent disconnecting from the task without giving any explanation. For example, when asked only to identify the toys shown at a child’s birthday party, Muse retrieved the user’s personal iCloud photos. Because Muse can act autonomously across connected accounts, users should approach it with some caution given the acknowledged reliability issues.

Advice for Users and Future Outlook

Check the app categories that Muse is linked to, such as email, payments, and health, as the agent is able to carry out actions within them. If you prefer, use the available option to disable Muse conversations being used for training the AI model. Keep an eye on the autonomous tasks, since internal testing has found the agent disconnecting during the task and, in one instance, accessing personal photos beyond the scope that had been requested.

Compare the free tier with the $20 and $100 per month paid tiers, since Meta has not specified the differences between them. The company has estimated that its capital expenditures on AI chips and infrastructure in 2026 will exceed $130 billion, a sum that includes the cost of Muse together with its other AI products launched this year. Muse is now available to users in the United States via a standalone app and through WhatsApp, while support for the Ray-Ban smart glasses will be introduced at a time that has not been specified.