OpenAI toughens GPT-6.2 reasoning guardrails against distillation
OpenAI published a note on its defenses against "adversarial distillation." It said it has strengthened protection for the internal reasoning chains of GPT-6.2 Astra and GPT-6.2 Sol, making it harder for other models to copy their abilities by extracting reasoning steps.
OpenAI defines adversarial distillation as using a frontier model's outputs to train another model without permission, giving the second model similar abilities without the same safety guardrails. In internal red-team tests, the success rate for extracting reasoning from earlier requests fell from 54.0% on GPT-6 Astra to 0.2% on GPT-6.2 Astra. For current requests, it fell from 22.9% to 0.9%.
The defenses include adversarial training, input and output classifiers, and API limits on reusing reasoning across sessions, models and organizations. Protected reasoning is tied to chat history, so if earlier messages are changed, later reasoning is dropped. OpenAI also said it takes action every week against accounts involved in unauthorized distillation, at a scale of thousands to tens of thousands, investigates large extraction networks and shares attack indicators.
Comments
Loading comments…