Meta Introduces Muse: A Personal AI Agent
Meta has unveiled Muse, a personal AI agent designed to facilitate digital task automation for users within a secure cloud environment. The company emphasises that security and privacy are “built in from the ground up.” Muse is now available for download on iOS and Android through an official app, as well as via the website Muse.ai. Users can also interact with Muse directly through WhatsApp, and Meta anticipates that users of its smart glasses will soon be able to engage with the agent as well.
While Muse can be tested free of charge, those interested in automating a wider array of digital tasks will need to subscribe to one of Meta’s AI subscription plans.
A New Player in the AI Landscape
Muse represents Meta’s latest effort to enter the competitive AI market, specifically against popular messaging-based agents like OpenClaw and Instinct. These platforms enable users to send messages that trigger various digital automations. Muse is a product of Meta’s Superintelligence Labs, a division established by CEO Mark Zuckerberg approximately a year ago to rival AI leaders such as OpenAI and Anthropic. The lab has attracted top researchers with enticing compensation packages, motivated by the belief that AI agents will revolutionise everyday internet navigation and Meta’s product offerings.
Previously known by the codename “Hatch,” Muse has undergone internal testing within Meta, where employees utilised it to autonomously manage third-party applications and browse the web on their behalf.
According to Meta, Muse can be set up effortlessly, designed with the user experience in mind to eliminate any learning curve. Users can issue natural language commands to the agent, which will autonomously handle tasks such as sending emails, booking travel arrangements, or even assisting in the sale of a vehicle.
Innovative Payment and Security Features
Muse also offers the capability to make purchases on behalf of users, leveraging a unique payment infrastructure developed in collaboration with Stripe. This payment tool, referred to as Link, generates a single-use card number, thereby eliminating the need for agents to input sensitive financial information online. Meta asserts that Muse is the first AI agent to be backed by Link’s purchase protections, ensuring commission-free refunds.
Though Meta is entering the personal assistant arena later than its competitors, it appears to be differentiating itself through a strong focus on security and privacy. The assistant utilises an architecture named Secure VM, which aims to isolate each user’s activities within a designated virtual machine, thereby keeping unreliable web data and integrations separate from the components that act on the user’s behalf.
Building Trust with Users
Trust is paramount for a personal AI agent, a challenge that Meta has faced over the years. In an effort to encourage users to try Muse and grant it access to various third-party applications and services, the company is working to assure them of its responsible data management practices. David Singleton, Vice President of Product Engineering at Meta’s Superintelligence Lab, explains, “If we are to create a product that can access numerous personal data sources, we must act responsibly. This system has been deliberately designed with that in mind.”
Singleton elaborates on the implementation of ‘Sentinel,’ a monitoring system that oversees all actions emanating from the virtual machine. It first checks whether the action is permitted under a policy already authorised by the user or system. If not, it prompts the user for approval before proceeding, ensuring that requests for confirmation are directed to the user without being filtered through the model, thus guarding against command injection attacks.
Enhancing Privacy Standards
Secure VM’s architecture is notable for its privacy considerations and sets a new industry benchmark for AI agents. Looking ahead, Meta plans to introduce Confidential VM, designed to run each virtual machine in a “trusted execution environment,” enabling users to manage their own access keys locally on their devices. This measure ensures that even Meta cannot access an individual user’s virtual machine.
Confidential VM is a product of Meta’s collaboration with Moxie Marlinspike, the creator of the end-to-end encrypted messaging app Signal, who has also developed the privacy-focused AI platform, Confer.
Meta has provided select security firms with access to the source code of Confidential VM for auditing and regular verification of its privacy assurances, and will publish the binaries of Confidential VM alongside a transparency log, allowing users to authenticate the integrity of their connection with Muse.
Commitment to Security Testing
Singleton asserts that Muse’s Secure VM has undergone extensive security testing, both through automated and human red teams at Meta, and is now included in the company’s public bug bounty programme. This initiative offers rewards of up to $300,000 for valid vulnerability discoveries, including up to $130,000 for successful command injection attacks affecting individual users.