We Built Safety into Muse (research.meta.ai)

Today, Meta unveiled Muse, a personal AI agent designed to seamlessly integrate into users' lives while enhancing safety and security. Operating from a dedicated cloud-based virtual machine (VM), Muse is engineered with a robust safety architecture that includes features like zero-shot tool calling, long context processing, and adaptive prompt handling. This careful engineering allows Muse to deploy sub-agents and tools autonomously, although its operation is closely monitored to prevent misuse, such as prompt injection attacks. The significance of Muse lies in its comprehensive safety measures, which include a separate authority, Sentinel, that governs all actions and interactions with external services, ensuring that sensitive data remains secure and inaccessible to the agent itself. Thorough testing through internal red teaming and a newly launched bug bounty program, with rewards of up to $300,000, underscores the commitment to identifying vulnerabilities preemptively. In doing so, Muse exemplifies a new standard in the AI/ML community, where user trust and safety are prioritized while harnessing the potential of personal superintelligence in everyday applications.