OpenAI updates GPT-6.2 Astra safeguards against adversarial distillation

openai.com9 min read

OpenAI details its strategy and updated safeguards for combatting adversarial distillation, a method bad actors use to extract frontier model capabilities and reasoning traces without authorization. The company reports that new request-level defenses and campaign detection significantly lowered the attack success rate for extracting protected reasoning in GPT-6.2 Astra compared to previous versions. OpenAI also highlights ongoing national security concerns related to unauthorized distillation by foreign actors and calls for stronger legislative and export controls. Open article →