Anthropic has privately pitched lawmakers a draft AI safety bill that would narrow when federal officials can stop a model’s release and override some state AI rules.