OpenAI: For the new stage of model capability development, we need to expand the scope of disclosure of misalignment events. Currently, we, and the AI industry more broadly, have yet to establish clear standards for reporting misalignment issues during training, evaluation, and deployment. It includes a variety of cases that are not traditional security incidents, but are useful for understanding AI behavior and predicting future risks. We are building a relevant framework that will be announced in the coming weeks; at the same time, we are collaborating with dozens of government regulators around the world on these issues.

OpenAI: For the new stage of model capability development, we need to expand the scope of disclosure of misalignment events. Currently, we, and the AI industry more broadly, have yet to establish clear standards for reporting misalignment issues during training, evaluation, and deployment. It includes a variety of cases that are not traditional security incidents, but are useful for understanding AI behavior and predicting future risks. We are building a relevant framework that will be announced in the coming weeks; at the same time, we are collaborating with dozens of government regulators around the world on these issues.