AI NewsModels & agentsAnnouncement
Apollo Research will embed six evaluators inside OpenAI for four months
Six Apollo Research staff will work inside OpenAI for four months. They will check whether its models try to hide bad behaviour.

Image: Apollo Research
Why it mattersAn outside team reading a lab's training data is new. The pilot shows what access an eval team needs, and what a lab is willing to give.
An AI lab checks its own models for bad behaviour. An outside eval company reads the finished model and looks for the same thing. Neither of them sees the middle: the training runs, the reasoning traces, the reward signals that shaped how the model behaves. Apollo Research says it will now sit in that middle.
The AI safety firm announced on 16 November that six of its researchers will work inside OpenAI for four months. Apollo Research calls it "a pilot towards embedded evaluations". The team will check whether OpenAI's models try to "subvert their own or their successors' alignment or control mechanisms".
What Apollo gets to see
OpenAI will give Apollo access to training code, training datasets, and reinforcement-learning rollouts. The rollouts include the model's own reasoning traces and the reward signals that scored them. Apollo can also query intermediate model checkpoints. All of the work happens on OpenAI hardware. This is more access than an outside evaluator usually has to a frontier lab.
What is left out
The access has limits that Apollo names in its own post. OpenAI can withhold information inside the agreed scope for security, technical, or intellectual-property reasons. Apollo can only speak to authorised OpenAI staff. A regular OpenAI employee who wants to raise something in the scope of the pilot cannot reach the Apollo team. OpenAI may redact sensitive material from the final reports. Apollo says the deal does not let OpenAI redact a bad finding on its own.
The four questions
The four months of work are shaped around four questions, all listed in the Apollo post. Does the training process reward subversion? Can evaluations draw subversion out of a model? Does subversion happen while a model is deployed inside OpenAI? Would the reasoning that led to subversion be detectable at all? Two public reports will come out at the end, covering the methods, the findings, and the limits.
Why this matters for the people building on OpenAI's models
A team building on OpenAI's API cannot run evaluations like these. The training data and the intermediate checkpoints are not in the API. So the only check they can run is at the output of the finished model. Apollo's pilot may find a subversion pattern that shows up earlier in training. A builder would not get that from any other source. The pilot is also honest about its limits. It names what access Apollo could not get, and what it could not say in public. When the reports come out, a builder can compare them against their own eval suite's checks on the finished model.
Source
- A Pilot Towards Embedded Evaluations, Apollo Research
- OpenAI will embed Apollo researchers to investigate AI subversion, TechCrunch
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.


