- Prior-request extraction dropped from 54.0% to 0.2%, current-request from 22.9% to 0.9%, aggregated across six attack models
- Defenses span adversarial training, input/output classifiers and restrictions on cross-session and cross-org reasoning reuse
- OpenAI acts weekly against thousands to tens of thousands of unauthorized distillation accounts and shares attack indicators with other providers
AI CURATED HOT STORY · 11/20/2026
OpenAI Hardens Reasoning Defenses, Cuts Extraction Attack Success to 0.2%
OpenAI has systematically disclosed a multi-layered defense framework against adversarial distillation, achieving significant improvements: the success rate of extracting protected reasoning from historical requests dropped from 54.0% to 0.2%, and for active requests from 22.9% to 0.9%. The defense covers frontier models including GPT-6.2 Astra and Sol, with some improvements backported to GPT-6 Astra.
The framework combines three layers: per-request adversarial training, input/output classifiers, and API-level reasoning isolation; cross-session and cross-organization reasoning reuse restrictions; and batch account bans for coordinated attacks. OpenAI takes enforcement action against thousands to tens of thousands of unauthorized distillation accounts weekly and shares attack indicators with other vendors. The official blog notes that full reasoning traces are more valuable for distillation than summaries on SWE-bench Verified and frontier cybersecurity tasks, so defenses must cover reasoning chain outputs.
This disclosure carries dual significance: adversarial distillation allows attackers to replicate frontier capabilities without safety guardrails, posing real abuse risks, and OpenAI's framework offers reusable cross-industry lessons. Meanwhile, Arena.ai's updated Alignment Index shows OpenAI sweeping the top seven spots, with Claude Fable 5.2 debuting eighth and Kimi K3.1 Pro the highest-ranked open-source model at ninth, underscoring that safety and alignment are becoming key competitive dimensions.
Hourly heat curve
24 fully observed hoursThe line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.
Coverage timeline
Total 5 reports · Latest first- IT之家 · 智能时代T2·78 pts
- The DecoderT2·78 pts
- Prior-request reasoning extraction ASR fell from 54.0% to 0.2%; active-request attacks from 22.9% to 0.9%
- Two-layer defense: per-request training, input/output filters and API blocks, plus mass account bans against coordinated campaigns
- Selected improvements are being backported to GPT-6 Astra, extending defenses to earlier models
- Hacker News AIT2·78 pts
- Adversarial distillation lets actors replicate frontier capabilities without safeguards, with cyber offense a prime abuse case
- OpenAI layers adversarial testing, training, classifiers and API protections to detect and disrupt it
- Hardening protected reasoning is central, and the lessons are meant to be reusable across the industry
- Arena.ai (@arena)T1·74 pts
- OpenAI sweeps the top seven spots, led by GPT-6.2 Sol
- Claude Fable 5.2 debuts at #8 while predecessor Opus 5.5 falls from #6 to #17
- Kimi K3.1 Pro is the top open model at #9, with open releases still trailing closed ones
- OpenAI 官方博客T1·78 pts
- Adversarial distillation lets attackers replicate frontier capabilities while stripping safeguards, creating real abuse risk
- OpenAI layers adversarial testing, training, classifiers, and API protections into a combined defense
- Full reasoning transcripts are more distillable than summaries, so defenses must cover reasoning outputs