- Chain-of-thought extraction fell from 54% to 0.2% on legacy requests and 22.9% to 0.9% on new ones
- OpenAI layers defenses and bans thousands to tens of thousands of unauthorized distillation accounts weekly
- Reasoning traces are now treated as core assets requiring active protection
AI CURATED HOT STORY · 11/21/2026
OpenAI Cuts Chain-of-Thought Extraction Success Rate to 0.2%
OpenAI has announced progress in anti-distillation protections for GPT-6.2 Astra and GPT-6.2 Sol. Black-box red-team testing shows that the success rate of extracting reasoning chains from old requests dropped from 54% to 0.2%, and for new requests from 22.9% to 0.9%. OpenAI also employs multi-layered defenses and bans thousands to tens of thousands of accounts weekly for unauthorized distillation, marking a substantive confrontation phase in protecting reasoning chain assets.
Technically, OpenAI's multi-layered defenses significantly reduce extraction success rates. The bans on unauthorized distillation accounts indicate active measures against model theft. These efforts collectively protect reasoning chain assets.
Industry-wide, reasoning chains are now viewed as core assets requiring protection. OpenAI's actions may prompt other frontier model vendors to follow suit, making anti-distillation technologies an industry standard. This could also intensify the cat-and-mouse game between model providers and distillation attempts, potentially driving further regulatory developments.
Hourly heat curve
25 fully observed hoursThe line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.
Coverage timeline
Total 2 reports · Latest first- IT之家 · IT业界T2·58 pts
- IT之家 · 智能时代T2·52 pts
- Tao Lin predicts AI will match or surpass top human experts in nearly all domains by late 2027
- He warns job displacement may unfold in waves, ultimately leaving almost everyone reliant on government support
- Mass unemployment by 2027 is highly unlikely due to the slow pace of robotics deployment