The risk threshold under Anthropic's draft AI safety bill must stem from a cyberattack, a model evading developer controls or going beyond human control, or the creation or release of a biological weapon