OpenAI details adversarial distillation defenses
Stronger protection of GPT-6.2 reasoning: extraction attacks fell from 54% success on GPT-6 Astra to 0.2%.
We’ve strengthened our defenses against protected reasoning extraction in GPT‑6.2 Astra and GPT‑6.2 Sol.
Official source
Combatting adversarial distillationopenai.com · published