PolicyNov 20, 2026, 5:00 AM EST

OpenAI details safeguards against adversarial distillation

OpenAI published an account of how it detects and disrupts adversarial distillation, the practice of training a second model on a frontier model's outputs so it picks up similar capabilities without the original's protections. The company says its defenses combine adversarial training, classifiers, API-level controls and investigations into coordinated extraction campaigns, with particular focus on stopping outsiders from recovering a model's internal reasoning traces. OpenAI reports that in internal automated red-teaming, the average attack success rate on protected reasoning fell from 54.0% against GPT-6 Astra to 0.2% against GPT-6.2 Astra, which carries the updated safeguards along with GPT-6.2 Sol. The write-up is public and describes the approach rather than shipping a new product.

Sources

Written by AIToolMaps from the source above and checked against it sentence by sentence.

Comments

No comments yet