Frontier Labs Harden Models Against Reasoning Extraction and Distillation
Frontier labs will treat their models' hidden reasoning and outputs as assets to defend against rival distillation. Extraction resistance will be trained into each release and reported as a headline metric, account-level enforcement will run at industrial scale, and labs will increasingly share threat intelligence with each other, making anti-distillation a recurring security and competitive category.
weakening · confidence 36 · Emerging (watchlist) · tracking since November 21, 2026 · updated November 22, 2026
Score history
Daily conviction score, 0 to 100. Higher means the thesis is more strongly corroborated.
Now 36 · -2 since Nov 21 · ranged 36 to 38
Why the conviction moved
- Nov 21Strengthened
OpenAI says GPT-6.2 Astra leaks protected reasoning in 0.2% of adaptive extraction attacks, down from 54% on GPT-6 Astra. The company says it now acts against thousands to tens of thousands of accounts a week and shares indicators of compromise with other model providers.
Source trail
Supporting · November 21, 2026
OpenAI says GPT-6.2 Astra leaks protected reasoning in 0.2% of adaptive extraction attacks, down from 54% on GPT-6 Astra
OpenAI says GPT-6.2 Astra leaks protected reasoning in 0.2% of adaptive extraction attacks, down from 54% on GPT-6 Astra. The company says it now acts against thousands to tens of thousands of accounts a week and shares indicators of compromise with other model providers.
OpenAI News
Unlock full source trail, score history, and daily updates.
Unlock TrendsTownsquare
Argue the thesis in Townsquare.