Disrupting a coordinated model-distillation campaign

OpenAI identified and disrupted a coordinated campaign in which operators systematically extracted protected reasoning from OpenAI's models through manipulated interactions, rather than through direct hacking or database breaches. Protected reasoning is the model's internal working process, and extracting it can reveal information hidden from final answers and help others reproduce the model's capabilities. This type of attack, called adversarial distillation, poses safety and national security risks because extracted reasoning could be used to train other models without the original safeguards, and at scale it can accelerate the transfer of advanced capabilities without equal investment in safety measures. OpenAI responded by banning fraudulent accounts, strengthening technical controls, and sharing findings with industry partners and government channels so other developers can defend against similar attacks.