OpenAI to host Apollo Research evaluators in pilot to probe model subversion
Apollo Research says OpenAI will embed six of its safety evaluators in a pilot to investigate whether models attempt to subvert alignment or control mechanisms; Apollo did not disclose a start date. The pilot grants Apollo limited access to training code, datasets, rollouts and checkpoints for agreed models and periods but allows OpenAI to withhold some information for technical, security or IP reasons and restricts researcher access to authorized staff and disclosure of certain evidence.