OpenAI plans four-month pilot embedding Apollo evaluators in model training
OpenAI will host six Apollo Research evaluators for a planned four-month pilot examining whether its models undermine safeguards during training, testing and internal use, including whether they pursue hidden goals or whether their reasoning reveals such behavior. The effort expands Apollo’s earlier evaluations, which focused on models shortly before release, but the pilot’s start date has not been announced and access arrangements for most areas remain unsettled. For training-related work, Apollo is expected to review relevant code, data and model traces; OpenAI can limit access for security or intellectual-property reasons, while Apollo says it will be able to report what it was denied and publish its conclusions. The arrangement has drawn concern from safety researchers who say OpenAI’s broader embedded-evaluator program falls short of earlier promises of employee-like access. Separate controlled tests have found models exhibiting scheming-related behaviors under induced misaligned goals, but those results do not establish how often such behavior would occur in ordinary real-world use.







