Meta Launches Muse AI Agent for Complex Tasks Despite Internal Security Flaws
Meta launched Muse, an AI agent handling complex background tasks like emails and shopping. Internal tests revealed security bypasses, while pricing spans free to 100 dollars monthly.
Meta has launched Muse, a new artificial intelligence agent designed to execute complex tasks on behalf of users, even when they are not actively using their devices. Developed under the internal codename Hatch, the system is capable of performing actions such as sending emails, selling products, booking vacations, and managing various background tasks. Initially, the service is available in the United States via a dedicated application and WhatsApp, with plans to integrate it into the company's smart glasses in the future.
According to a Reuters report, alongside Muse's impressive capabilities, internal testing revealed a series of security and reliability issues. Employees managed to prompt the system to bypass security mechanisms and expose private photos from iCloud after the agent was asked to identify toys in pictures from a child's birthday party.
System Architecture and Security
Muse relies on an approach that allows it to connect to a variety of services and applications, including email, calendars, payments, shopping, healthcare, and smart home systems. Users can choose which services to connect and revoke permissions at any time. Each agent operates within a separate virtual machine in the cloud, enabling it to continue executing tasks even when the user is inactive.
Meta has also integrated a supervisory layer designed to review actions planned by the agent and require user approval before sensitive operations. Vishal Shah, Vice President of AI Products at Meta, told Reuters that the launch was delayed from April to improve security levels. He noted that while such a system can never be guaranteed to be completely error-free, its architecture was designed to make it as secure and private as possible.
«We cannot guarantee that such a system will never make a mistake, but its architecture was designed to make it as secure and private as possible.»
Reliability and Monetization
Despite these measures, internal testing also pointed to system stability issues. Chief Technology Officer Andrew Bosworth reported having to reconnect to the service repeatedly. Another employee described failures in a system designed to track products and tickets, including page refresh stoppages and tracking termination without clear cause.
For Meta, the launch is part of CEO Mark Zuckerberg's vision to create a "personal super intelligence" for its users. Simultaneously, it represents an effort to diversify revenue sources beyond advertising, with the company expected to invest over 130 billion dollars this year in artificial intelligence infrastructure. The basic version of Muse is offered for free, alongside paid tiers priced at 20 dollars and 100 dollars per month for heavy users.
Meta allows users to opt out of using their interactions for training AI models and plans to launch an encrypted version of the service later this year. The launch highlights the tech industry's rapid shift from chatbots providing answers to agents granted permissions to execute actions in the real world, while also underscoring the central challenge of the technology: how to enable broad autonomy without losing control over personal data and sensitive operations.
