Apollo Research will station six evaluators in OpenAI's sabotage pilot
Apollo Research announced on November 16, 2026 that six evaluators will work inside OpenAI for four months to investigate whether models attempt to undermine alignment or control mechanisms during training, evaluation, and internal deployment. Previously, evaluations were conducted only after post-training was complete and shortly before public release; Apollo believes that timing is insufficient to test behaviors such as scheming, and this embedded arrangement gives evaluators employee-like access. OpenAI commits to providing training code, datasets, reinforcement learning training records, and the ability to query intermediate checkpoints for training incentive evaluations; it may withhold materials for technical, safety, or intellectual property reasons, but must not withhold them solely because the information is unfavorable. Apollo plans to publish two public reports; OpenAI may request removal of confidential information but cannot alter evaluation conclusions, and Apollo may also publicly state the access it received and the restrictions it faced, and says this is an ambitious first step, not a complete embedded evaluation.
The original source discloses in detail the scope of access for the evaluation, the right to redact reports, and the meta-transparency provisions, so readers can weigh the significance of this embedded evaluation's conclusions accordingly.