What happened
Meta Superintelligence Labs published the Muse Spark 1.1 Evaluation Report (dated July 9, 2026) alongside the model's public release, detailing its safety profile under Meta's Advanced AI Scaling Framework across Chemical & Biological, Cybersecurity, and Loss of Control risk domains. The headline disclosure: 'when evaluated without mitigations applied, we cannot rule out Muse Spark 1.1's capabilities meeting [the high-risk] threshold in both the Chemical & Biological and Cybersecurity domains' — though Meta states multi-layered mitigations reduce residual risk to 'moderate or lower' in all domains prior to release. The 100+ page report also covers adversarial robustness (AgentHarm, AgentDojo, third-party red-teaming via GraySwan ART), model behavior (scheming, sycophancy, evaluation awareness, situational awareness of training/deployment context), and general capability benchmarks, reflecting an expanded scope given the model's new agentic API affordances (tool/function calling).
Why it matters
This is a frontier lab's first public acknowledgment that a shipped model's pre-mitigation capabilities could not be ruled out from crossing catastrophic-risk thresholds in CBRN and cybersecurity domains — a disclosure CISOs and AI governance leads should factor into vendor risk assessments for agentic API deployments.
Action needed
Have technical/security teams review the mitigation and red-teaming methodology sections to benchmark internal AI risk evaluation practices against Meta's disclosed framework.