Key takeaways
- OpenAI is embedding evaluators from Apollo Research, Goodfire, METR and Transluce in a pilot announced November 16, 2026.
- Apollo's six staff get training code, datasets and mid-training checkpoints for four months to look for models undermining their own safety controls.
- OpenAI can ask to redact confidential evidence from Apollo's reports, but not its verdicts.
OpenAI will let four outside safety groups work inside the company to test its models, starting with a pilot announced on November 16, 2026. Apollo Research said six of its staff will spend four months checking whether OpenAI's models try to undermine their own safety controls, or those of the models that come after them. Goodfire, METR and Transluce are the other three, TechCrunch reported.
The difference from normal outside testing is access. Labs usually hand evaluators a finished model through an API. Apollo instead gets the relevant training code and datasets, reinforcement learning runs with their reasoning traces and rewards, and the ability to query checkpoints from partway through training. It will look at four questions: whether training rewards subversion, whether tests can draw it out, whether it shows up when OpenAI uses its own models internally, and whether the reasoning behind it can be spotted.
The other groups have narrower jobs, according to TechCrunch. Goodfire brings interpretability tools to monitor training runs. METR runs its own tests for dangerous behavior and checks whether OpenAI's monitoring would catch it. Transluce reviews whether OpenAI's post-training safety evaluations are good enough.
Apollo calls the deal a pilot that does not yet meet its own bar. OpenAI can ask to redact confidential evidence from Apollo's two planned reports, though not Apollo's verdicts, and access for three of the four questions is still being negotiated. Apollo works on OpenAI hardware and has no access to internal channels such as Slack by default. Apollo says the arrangement is "an ambitious first step towards deeper access and disclosure rights."
For readers who use OpenAI's models, the reports are the part to wait for. They are the first time an outside group will say in public what it found inside a frontier lab's training process. Our explainer on AI red teaming covers how labs test their own models today.

