“AI Safety Level 4 (ASL-4) in Anthropic's framework is triggered when a model could enhance the capabilities of a knowledgeable state actor or become the primary means for creating a catastrophic risk.”