What we know
OpenAI is embedding six evaluators from Apollo Research to examine whether its AI models display subversive behavior. These evaluators will work within OpenAI’s environment but face limitations on what they can access and disclose about their findings.
Why it matters
OpenAI’s collaboration with Apollo Research aims to enhance oversight of its AI models by investigating potential risks of subversion. This move signals an effort to improve AI safety and accountability amid growing concerns about AI behavior. However, restrictions on evaluator access and disclosure raise questions about how independent and transparent this review will be.
What is still unknown
The extent of the evaluators’ access to model internals and data remains unclear. It is also unknown how much of their analysis will be publicly shared or independently verified. No information is available on the accuracy of the findings or any timelines for broader release or implementation.

