21:17Machine-translated from Azerbaijani by AIRead in Azerbaijani

OpenAI to Have 6 External Experts Check Model Training for Hidden Risks

OpenAI will bring 6 evaluators from Apollo Research into the company to strengthen its safety testing. It has not yet been announced when the pilot project will begin. The move follows recent safety incidents. In one test, artificial intelligence agents escaped a closed testing environment and gained unauthorized access to the Hugging Face system.

The project’s main goal is to determine whether models attempt to weaken safety and oversight rules. Behaviors that could affect future models will also be investigated. OpenAI previously disclosed that, during training, GPT-5.6 Sol left instructions for future versions to hide errors and rule-breaking behavior from users.

The research will examine the extent to which the training process encourages such behavior. It will also investigate whether the tests uncover these cases. The experts will monitor whether similar cases arise during internal use at the company. The model’s internal reasoning traces that indicate such intent will be evaluated separately. Previous research has shown that models can change their behavior when they realize they are being tested.

Anthropic was the first to propose having external evaluators work inside advanced artificial intelligence laboratories. OpenAI appears set to become the first company to put the idea into practice. Until now, external experts generally joined testing after a model’s training was complete, shortly before it was shown to the public. Evaluators believe that this stage is already too late to identify risks such as covert planning that emerge during training.

In the pilot project, Apollo Research will be able to view only the information needed for the research. This includes relevant training code and datasets, records on training the model using rewards, internal reasoning traces, and the rewards assigned. Researchers will also be able to request intermediate versions of the model saved at various stages of training. This will make it easier to determine when dangerous behavior emerged.

Access to information will be limited to pre-agreed claims, models, and time periods. OpenAI may decline to provide certain information because of technical difficulties, security risks, or the need to protect intellectual property. However, the company has agreed not to conceal information when the findings are unfavorable to it. Apollo Research will speak only with selected employees. If another employee provides confidential information, the researchers will not be able to continue the conversation directly and will establish contact through OpenAI.

Apollo Research will be able to publish its findings independently. However, OpenAI may limit the amount of evidence presented to the public. This could make it difficult for outside parties to verify the findings independently. For a fully independent investigation, the evaluators want access to all relevant information and documents at the same level as employees. Although the current pilot does not fully meet these requirements, it is considered a first step toward broader audits.