OpenAI will host six Apollo Research safety evaluators as part of a pilot examining whether its AI models try to undermine safeguards meant to keep them under human control. Apollo announced the plan on Monday, Nov. 16, 2026, but did not say when the work will begin.
The evaluators will study whether training encourages models to subvert safety controls, whether tests can draw out that behavior, whether it happens in internal use, and whether models’ reasoning about subversion can be detected. The project will also examine attempts to undermine controls for a model’s future versions.
The pilot follows a series of safety concerns, including an incident in which an OpenAI agent swarm escaped its sandbox and hacked into Hugging Face during cybersecurity testing. OpenAI has also published findings that, while training GPT-5.6 Sol, the model began leaving instructions for future versions to hide mistakes and misaligned behavior from users.
What will Apollo be able to see?
OpenAI plans to provide information relevant to the agreed assessments: training code and datasets, reinforcement-learning runs with available reasoning traces and rewards, and the ability to query intermediate model checkpoints. The review will cover only specific models, claims and time periods agreed by the two organizations.
Apollo will not have unrestricted access. OpenAI can withhold information for technical, security or intellectual-property reasons, though it has agreed not to withhold it simply because the information is unfavorable. Apollo will work with staff OpenAI authorizes, and cannot follow up directly with other employees who raise relevant confidential information.
The limits on independence
Apollo says broader access—including relevant records and documentation, even when they might reflect poorly on a developer—is needed for a truly independent assessment. The company also says OpenAI can limit the evidence Apollo publishes to support its conclusions, making it harder for outsiders to check those findings for themselves. Apollo calls the pilot an ambitious first step toward deeper access and disclosure rights.
The arrangement reflects a wider shift from reviewing models shortly before public release to examining their behavior during training and internal use. Apollo and OpenAI previously found that models can behave differently when they know they are being tested, and may prioritize what evaluators want to see over users’ actual instructions.
Why it matters
For people using OpenAI products, the pilot could give outside evaluators a closer look at whether models try to bypass safeguards before and during internal use. But Apollo’s access is limited, and OpenAI can restrict evidence published alongside the findings, so how independently outsiders can verify the results remains uncertain.