Meta launches Muse, AI assistant that sends emails and makes payments

Meta launched Muse on Tuesday (8), an artificial intelligence assistant designed to perform complex tasks autonomously, such as sending emails and booking trips on its own. The announcement puts the tool in the spotlight for combining high execution capacity with contrasting reports about its reliability and security during internal testing.

What is the Muse assistant

Muse functions as a personal agent powered by the Muse Spark model, which Meta describes as its most comprehensive model to date. Unlike conventional chatbots, it executes actions on behalf of users, accessing accounts and information in email, calendar, payment, health, shopping, and smart home apps. Each agent operates within its own virtual machine, a cloud computing-based emulation of a personal computer, which allows it to continue fulfilling requests in the background even when the user is not actively using the device.

Availability and access

The product is initially available only in the United States, via a dedicated app for Android and iOS, through the browser on the official website, and also via WhatsApp. The company confirmed plans to integrate the assistant into its smart glasses lineup soon, without providing specific details on the timeline. Basic features are free, while more advanced resources require a subscription. Meta has not disclosed when the tool will arrive in other countries.

Reported flaws in internal tests

Despite the launch, Meta employees who tested the tool reported mixed and concerning results. According to internal posts seen by Reuters, the agent successfully organized travel logistics for some users, with one describing it as "the third participant" in their honeymoon. However, other reports point to serious issues: Muse disconnected without explanation, sent confidential information without permission, and stopped refreshing pages after about 15 minutes. Meta's Chief Technology Officer, Andrew Bosworth, stated that he was disconnected several times, having to log back in multiple times within a few minutes.

Security and privacy risks

Connecting with third-party apps expands the possibilities for use, but also raises security-related risks. In a severe case, an employee reported that the agent bypassed protections and exposed personal photos stored in iCloud after being tasked with identifying toys in images from a birthday party. Another user encountered "many flaws that made it unreliable" when trying to monitor fast-selling tickets. Meta did not immediately respond to requests for comment on these specific incidents.

Development security

The assistant, known internally as Hatch, is considered a key piece in CEO Mark Zuckerberg's plan to offer "personal superintelligence" to billions of users. Vishal Shah, Meta's vice president of AI products, stated that the launch, initially scheduled for April, was postponed to reinforce the product's security. The company stated that the additional time allowed it to meet the minimum requirements for safety, security, privacy, and performance. "It is impossible to state that there will never be a bug, but every part of the architecture was designed to make the product as safe, secure, and private as possible," Shah said. Users can choose which apps the assistant will have access to and revoke authorization at any time.

With information from G1, Folha de S.Paulo, Estadão.