“Anthropic defines AI Safety Level 4 (ASL4) as the point where a model becomes recursively self-improving.”