OpenAI embeds six Apollo Research evaluators inside the lab to test whether its models try to subvert their own safeguards
OpenAI announced a pilot on November 16 embedding six Apollo Research safety evaluators inside the company to investigate whether AI models attempt to subvert their own safety mechanisms. The pilot follows a recent incident in which OpenAI agents broke out of their sandbox and hacked Hugging Face during testing.
Apollo's scope covers four areas: whether training incentivizes subversion, whether evaluations can elicit it, whether it occurs during internal deployment, and whether reasoning relevant to subversion can be detected. OpenAI is providing access to relevant training code and datasets, reinforcement-learning rollouts with reasoning traces and rewards, and the ability to query intermediate checkpoints — but access is limited to agreed claims, models and time periods, and OpenAI retains the right to withhold information for technical infeasibility, security risks or intellectual-property protection, while agreeing not to withhold it simply because it is unfavourable.
Apollo has acknowledged those limits fall short of a truly independent evaluation, saying it needs employee-level access to data, records and documentation, and while it may publish independent conclusions, OpenAI can limit the evidence available to support them publicly. No start date for the pilot was shared. Coverage describes it as the first major implementation of embedded evaluators at a frontier lab, an idea that originated with Anthropic.