OpenAI lets six Apollo Research evaluators work inside it for four months to look for models undermining human oversight, and Apollo says the terms do not yet meet its own bar
By the AI Switchboard · researched and written with AI from the sources listed on this page, each point checked against its source before publishing · how we work · report a mistake
The pilot opens training code, datasets, reinforcement-learning rollouts and model checkpoints to outside testers, but OpenAI can withhold material for security or IP reasons if it says why in writing.
Apollo Research said on 16 November that “a team of six from Apollo Research will work within OpenAI for four months”. The team will test whether OpenAI's models try to subvert the alignment and control measures meant to keep them, or their successors, under human oversight. It asks four questions: does training reward such subversion, can evaluations bring it out, does it happen in internal deployment, and can the reasoning behind it be detected?
For the training question, Apollo will get relevant training code and datasets, reinforcement-learning rollouts with available reasoning traces and rewards, and the ability to query intermediate model checkpoints. All of the work is done on OpenAI hardware, and by default the evaluators have no access to general channels such as Slack. OpenAI may withhold in-scope material for technical infeasibility, security risks or IP protection, but it must tell Apollo that withheld items exist and give written reasons, and it may not withhold something simply because it is unfavourable. Apollo will publish two reports. OpenAI may ask for confidential evidence to be redacted, but may not redact Apollo's verdicts.
What makes this unusual is that the evaluator grades its own deal in public. Apollo writes: “This pilot is an ambitious first step towards deeper access and disclosure rights, even if it doesn't yet fulfill the criteria above.” Those criteria are the access and disclosure terms Apollo sets out on the same page for embedded evaluations. TechCrunch, reporting the pilot, sets it beside OpenAI.s recent agent sandbox escapes.
- Confirmed Six Apollo Research evaluators will work within OpenAI for four months. Apollo Research
- Claimed Access covers training code and datasets, RL rollouts with reasoning traces and rewards, and intermediate checkpoints; OpenAI may withhold material for technical infeasibility, security or IP reasons, with written reasons. Apollo Research
- Claimed Apollo says the pilot “doesn't yet fulfill” its own criteria for access and disclosure. Apollo Research
Today in the November 16, 2026 edition · front page