Muse Spark Safety & Preparedness Report

Meta has released Muse Spark, its latest large language model, after internal safety evaluations under its Advanced AI Scaling Framework identified elevated risks in chemical and biological capabilities that were subsequently reduced through mitigations [1][2]. The model's preparedness was assessed across Chemical and Biological, Cybersecurity, and Loss of Control risk domains. Evaluations conducted before safeguards were applied found that Muse Spark's Chemical and Biological capabilities were likely reaching the "high risk" category under the company's Advanced AI Scaling Framework [1][2]. Meta then implemented a multi-layered set of mitigations. Following those changes, the company reports that Muse Spark demonstrates state-of-the-art refusal across a range of benchmarks related to hazardous workflows in chemistry and biology [1][2]. Meta concluded that deploying Muse Spark as the underlying model of Meta AI presents acceptable levels of residual risk under its framework [1][2]. Meta Platforms, headquartered in Menlo Park, California, owns Facebook, Instagram, WhatsApp, and Threads, and was the world's third-largest spender on research and development as of 2022, with R&D expenses totaling US$35.3 billion [8]. The company's investment in AI safety infrastructure was underscored in June 2025, when Meta spent more than $14 billion to acquire a 49% non-voting stake in Scale AI, a firm that operates an LLM Red Team conducting human adversarial testing to identify vulnerabilities and safety risks in AI models [9]. The release of Muse Spark comes as the broader industry grapples with the dual-use potential of advanced AI. The historical context for biological risk is stark: smallpox, an infectious disease caused by the Variola virus, is estimated to have killed up to 300 million people in the 20th century before its eradication was certified by the World Health Organization in 1980 [4]. The United States and Russia still retain samples of the virus in laboratories, a fact that has sparked ongoing debates over safety [4]. Meta's report on Muse Spark does not detail the specific hazardous workflows for which the model's refusal mechanisms were improved, but the company's public documentation frames the residual risks as acceptable for deployment [1][2].