← Reports and Research

Research report Independent evaluator Apollo Research

Topics: Evaluation and testing, Monitoring and control, Alignment research

Apollo Research plans a four-month check of OpenAI models from the inside

In short

Apollo Research, an AI evaluation group, says six of its staff will work inside OpenAI for four months. They will check whether OpenAI's models try to undermine their own safety controls, or those of later models. The post is an announcement of a plan. It contains no test results.

What happened

Apollo Research published the announcement on 16 November 2026. It calls the project a pilot toward "embedded evaluations," meaning outside evaluators work inside a developer's company with employee-like access.

Apollo says its earlier evaluations of frontier models (the most advanced AI systems) took place after training and shortly before public release. It says that is not enough for testing behaviours like scheming, which it uses to mean a model secretly working against its intended rules. It also says that approach cannot cover internal use of models that are not public.

The pilot will look at four questions:

  • Whether training rewards a model for undermining safety controls.
  • Whether evaluations can bring out that behaviour.
  • Whether it happens when OpenAI uses models internally.
  • Whether reasoning linked to it would be detectable.

Apollo says OpenAI has committed to access for the first question. That includes training code and datasets, records of reinforcement-learning runs (a method where models are trained with rewards), and the ability to query intermediate model versions. Access for the other three questions is still to be agreed.

Apollo says its access has limits. OpenAI may withhold some in-scope material for technical, security or intellectual property reasons. It must say that items were withheld and give written reasons. It may not withhold material only because it is unfavourable. All work will be done on OpenAI hardware, and OpenAI will choose which staff Apollo can speak to.

Apollo plans two public reports. The first covers training incentives. The second covers evaluations, internal use and detectability of the reasoning. OpenAI may ask for redactions of confidential evidence, including for sensitive intellectual property. It may not redact Apollo's verdict on each question. Claims that rest on staff conversations need formal confirmation from OpenAI before they can be published.

What was done about it

Nothing was being corrected here, because no problem was reported. The response is the arrangement itself. Apollo says OpenAI agreed that Apollo can publicly describe the conditions and limits of its work. That includes what was withheld, whether anything was removed from a report, and any pressure Apollo felt. Apollo says this "meta-transparency" meets its minimum standard. It also says the pilot does not yet meet its full bar for access and disclosure.

Why it matters

Apollo says that without meta-transparency, a report that finds no problems can look reassuring even when it only reflects limits on what the evaluator could see or share. It says the OpenAI agreement can serve as a public benchmark for other evaluations. Apollo says it hopes to run fuller embedded evaluations with several developers in future. Readers can find the full details in the original report.

This is a summary in everyday words, written by AI Misalignment Watch from the source above. If you spot a mistake, the original always wins.