“Anthropic advocates for a policy requiring all AI companies to conduct and publicly disclose the results of safety tests designed to elicit worst-case model behaviors.”