Anthropic has proposed embedding third-party safety evaluators inside frontier AI companies, and OpenAI has backed a similar arrangement, but critics say the outside reviewers would lack power to stop a model from being trained or released.