The law requires AI companies to address “catastrophic risks” from their models and mandates third-party audits of safety issues.