In OpenAI's internal adaptive red-teaming evaluation, the average attack success rate for recovering protected reasoning fell from 54.0% against GPT-6 Astra to 0.2% against GPT-6.2 Astra. The company described safeguards against adversarial distillation, where actors train models on a frontier model's outputs without inheriting its protections. OpenAI said its study found that protecting reasoning traces can limit cyber capability transfer, and outlined request-level and campaign-level defenses.
No score is assigned. Sources and their independence are shown in the citation chain below.