← Back to the wire

Combatting adversarial distillation

AnnouncementPolicyNov 20, 2026

In OpenAI's internal adaptive red-teaming evaluation, the average attack success rate for recovering protected reasoning fell from 54.0% against GPT-6 Astra to 0.2% against GPT-6.2 Astra. The company described safeguards against adversarial distillation, where actors train models on a frontier model's outputs without inheriting its protections. OpenAI said its study found that protecting reasoning traces can limit cyber capability transfer, and outlined request-level and campaign-level defenses.

Receipt № 29961 source · awaiting confirmation ◐

Evidence

1source· awaiting independent confirmation

No score is assigned. Sources and their independence are shown in the citation chain below.

Citation chain · 1 source

01high
openai.comOfficialhttps://openai.com/index/combatting-adversarial-distillation/This source is OpenAI's own technical write-up, giving an insider account of its distillation defenses, reasoning-trace protections, and internal capability-transfer study rather than outside reporting on them.
PRIMARY
OpenAICompany
Canonical: https://openai.com/index/combatting-adversarial-distillation/