OpenAI details new defenses against adversarial distillation of its models
OpenAI published an account of how it detects and blocks 'adversarial distillation,' where outside actors use a frontier model's outputs to train copycat models without the original's safety safeguards. The company says it combines adversarial testing, model training, classifiers, API-level monitoring and investigations to catch such attempts, and has specifically hardened protections around models' internal chain-of-thought reasoning. In internal red-team testing, OpenAI reports the success rate of extracting protected reasoning fell from 54.0% on GPT-6 Astra to 0.2% on the updated GPT-6.2 Astra.
GoKawiil's interpretation of the reporting above, not reported fact.
The disclosure suggests OpenAI views leaked chain-of-thought reasoning as a serious proliferation risk, since those internal traces could let attackers replicate a model's capabilities, including dangerous ones like cyberattack assistance, while stripping out the safety guardrails. By aligning distillation defenses with its biology and cybersecurity safeguards, OpenAI appears to be treating model-extraction attacks as a comparable tier of risk, which could push other AI labs to adopt similar protections for their own reasoning models.
- OpenAI says adversarial distillation lets bad actors copy a model's capabilities while stripping its safety safeguards.
- Protecting the model's internal chain-of-thought reasoning is now a key defensive focus, since it can reveal more than final answers.
- OpenAI reports attack success rates against protected reasoning dropped from 54.0% to 0.2% after upgrading from GPT-6 Astra to GPT-6.2 Astra.
Source: openai.com, 2026-11-20
Published there as: “Combatting Adversarial Distillation”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.