Meta launches Muse AI assistant with security-focused architecture
Just the facts

Meta launches Muse AI assistant with security-focused architecture

Summary

Meta has released Muse, a personal AI agent for iOS, Android and web, featuring a Secure VM environment and optional Confidential VM to protect user data, with free access and paid subscription tiers for extensive automation.

Meta announced the rollout of Muse, a personal AI assistant that users can message to automate tasks such as emailing, travel booking and online purchases. The service is available today through a dedicated app for iOS and Android, a web portal at muse.ai, and via WhatsApp, with plans to integrate it into Meta’s upcoming AI glasses. While basic use is free, advanced automation requires a subscription to Meta’s AI plans.

Developed by Meta Superintelligence Labs, the unit created by CEO Mark Zuckerberg to compete with other AI agents, Muse is built on a Secure Virtual Machine (Secure VM) architecture that isolates each user’s activity from external data and integrations. The system includes a “Sentinel” component that monitors outbound actions and prompts users for approval when needed, aiming to prevent unauthorized data access and attacks such as prompt injection.

Muse can also complete purchases using Stripe’s Link service, which generates single-use card numbers and offers purchase protections, including fee-free returns. Meta says the Secure VM does not allow the company to access user data, though it acknowledges that technical access would be possible under policy constraints; users may opt out of data use for model training.

Future enhancements will include a Confidential VM, a trusted execution environment where users control encryption keys locally, preventing even Meta from accessing the VM. The Confidential VM design is being reviewed with security expert Moxie Marlinspike, and Meta will provide source code, binaries and a transparency log for independent audits. The company has also expanded its bug bounty program, offering rewards up to $300,000 for valid vulnerabilities, including up to $130,000 for successful prompt-injection attacks affecting a single user.

"We know it’s really important, if we’re going to build a product like this that can access a lot of sources of personal data, that we’re really responsible with that, so we’ve designed this system very deliberately," said David Singleton, vice president of engineering for consumer products at Meta Superintelligence Labs.

Source

WIRED

Keep reading — for free

Create a free account to follow the news. No card required.