Meta announced an updated Advanced AI Scaling Framework and a forthcoming Safety & Preparedness Report for its Muse Spark model, emphasizing stronger safety, reliability, and user protections. The new framework expands on the company’s earlier Frontier AI Framework by broadening the range of risks assessed, including chemical, biological, cybersecurity, and loss-of-control scenarios. It also introduces stricter deployment standards and transparency measures through detailed reporting on risk assessments and evaluation results.

The Safety & Preparedness Report for Muse Spark highlights extensive pre-deployment testing across thousands of scenarios to identify weaknesses and ensure safeguards function effectively. Meta evaluated the model’s reasoning capabilities, potential for autonomous behavior, and adherence to safety policies covering violence prevention, child safety, and ideological balance. The company reports that Muse Spark demonstrated strong safeguards and minimal ideological bias.

Meta stated that safety mechanisms are integrated throughout the model’s lifecycle, from data filtering to product-level guardrails. The company’s evolving approach trains models to understand the reasoning behind safety principles, enabling them to handle new situations more effectively while maintaining human oversight and continuous monitoring.