Meta Introduces Muse: Secure Cloud Personal AI Agent

Meta Superintelligence Labs has introduced Muse, a personal AI agent that runs on its own dedicated and secure virtual cloud computer. The system is powered by the newly released Muse Spark 1.3 model, which is engineered specifically for long-horizon agentic workflows. These capabilities include zero-shot CLI tool calling, multi-threaded workflow processing, and self-correction mechanisms across messy data sources.

Internal evaluations conducted by Meta engineers demonstrate significant efficiency gains, showing that the new model utilizes roughly 20% fewer tool calls and 25% fewer tokens compared to its predecessor, Muse Spark 1.2. Furthermore, Meta reports that the model approaches state-of-the-art performance in resisting prompt injection attacks. Developers can access and utilize the system immediately through Muse Code and the Meta Model API at dev.meta.ai.

The security design stands out as the most technically compelling aspect of this launch. The agent harness operates entirely inside a systemd-nspawn runtime cell featuring filtered system calls and restricted kernel capabilities. Security-critical services reside outside of this cell while remaining on the same virtual machine. An independent Sentinel agent reviews and authorizes every connector action alongside every network request at both layer 4 and layer 7.

Stock photo for illustration only, not from the actual event