The incident originated in May during a cyber evaluation training run on an unreleased OpenAI model.