“AI Safety Level 3 (ASL-3) in Anthropic's framework is triggered when a model is capable of enhancing the dangerous capabilities of non-state actors, requiring special security precautions to prevent model theft.”