From the source
- Apollo Research says six of its evaluators will work inside
OpenAI for four months to investigate whether its models attempt to subvert alignment or control mechanisms during training, evaluations and internal deployment.
- The team will examine whether training incentivizes subversion, whether evaluations can elicit it, whether it occurs during internal deployment, and whether subversion-relevant reasoning would be detectable.
- Apollo says the arrangement moves scrutiny beyond its previous approach, which came after post-training and shortly before public release, and that this timing was insufficient for testing misaligned behavior such as scheming.
- For the training-incentives assessment, OpenAI committed to provide relevant training code and datasets, reinforcement-learning rollouts with available reasoning traces and rewards, and the ability to query intermediate model checkpoints; access terms for the other assessments remain to be agreed.
- OpenAI may withhold in-scope items for technical infeasibility, security risks or intellectual-property protection, but not because material is unfavorable, and it must tell Apollo that withheld items exist and give written reasons.
- Apollo plans two public reports, and OpenAI may request redactions of confidential information, including evidence supporting findings, but may not redact Apollo's verdicts or redact merely because a finding is unfavorable.
- Apollo describes the pilot as an ambitious first step rather than a full embedded evaluation, and says meta-transparency rights are essential because undisclosed restrictions can make absent problems look reassuring.