OpenAI hardens GPT-6.2 against distillation of protected reasoning
Original title: Combatting adversarial distillation
In OpenAI's internal red-team, recovering protected reasoning fell from 54.0% on GPT-6 Astra to 0.2% on GPT-6.2 Astra.
OpenAI says unauthorized distillation can copy frontier cyber and biology capabilities into a model that lacks the original safeguards. It reports stronger protections for exact chain-of-thought on GPT-6.2 Astra and GPT-6.2 Sol, combining training, classifiers, API controls, and disruption of coordinated extraction. An internal study found students trained on exact reasoning traces gained more cyber skill than those trained on user-visible summaries.
Unconfirmed: The 54.0% and 0.2% rates are from OpenAI's own internal evaluation, not an independent test.
- Internal adaptive red-team success fell from 54.0% to 0.2% on protected reasoning.
- Safeguards were strengthened on GPT-6.2 Astra and GPT-6.2 Sol.
- OpenAI says the controls now align more closely with biology and cybersecurity safeguards.
- Students were compared after supervised fine-tuning and light reinforcement learning.
- Organizations
- OpenAI
- AI models
- GPT-6.2 AstraGPT-6.2 SolGPT-6 Astra
From the original article
12 min read at openai.comOpenAI’s layered approach to adversarial distillation combines model training, safeguards, detection, and enforcement to protect against capability extraction.
openai.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.