Anthropic's draft AI safety bill requires frontier AI developers to test models, be transparent about findings, and undergo independent evaluations