Physix Frontier · News Briefing Card (IT Home · Nov 21, 2026)
OpenAI: GPT-6.2 Astra reasoning-chain extraction rate falls to 0.2%
KEY FACTS
- OpenAI published an announcement on November 20 detailing defenses against adversarial distillation.
- Reasoning-chain protection has been strengthened for GPT-6.2 Astra and GPT-6.2 Sol.
- In black-box red-team testing, the reasoning extraction success rate for historical requests fell from 54.0% to 0.2%.
- The reasoning extraction success rate for current requests fell from 22.9% to 0.9%, aggregating data from 6 attack models.
- OpenAI takes action every week against accounts engaged in unauthorized distillation, at a scale of thousands to tens of thousands.
KEY DATA
54.0%Historical-request reasoning extraction success rate (old mechanism)
0.2%Historical-request reasoning extraction success rate (new mechanism)
22.9%Current-request reasoning extraction success rate (old mechanism)
0.9%Current-request reasoning extraction success rate (new mechanism)
PHYSIX OBSERVATION
Pushing reasoning-chain extraction down to 0.2% shows that leading vendors have begun treating the reasoning process as a core asset to defend. For the industry, the distillation bar is raised and the path of replicating capabilities through wrapper shells narrows; for users, model capabilities are harder to copy at low cost, but restrictions on API reasoning reuse may also affect the multi-turn development experience.
Source: IT Home report
Physix Frontier