OpenAI cuts reasoning-extraction attack success rate from 54% to 0.2%
The Decoder · rss · 2026-11-21
Per The Decoder, OpenAI has significantly hardened safeguards against extracting internal reasoning from GPT-6.2 Astra and GPT-6.2 Sol.
- In adaptive red-teaming tests, the success rate of protected-reasoning extraction attacks dropped from 54% to 0.2%
- The company suspends thousands to tens of thousands of accounts weekly over coordinated distillation campaigns.
More from Models
- MaxxBench test: skip Qwen for coding, Muse Ember beats DeepSeek on price-performance — shuchaobi · 2026-11-21
- Mathematician flags concrete error on page 444 of Claude's claimed Riemann proof — _onionesque · 2026-11-21
- Claude's claimed Riemann work fails an Astra audit, follow-up reveals — basedjensen · 2026-11-21
- METR Time Horizons 2.0 returns: 3.6-month doubling time, GPT-6 Astra hits 40.3 hours — scaling01 · 2026-11-21
- Google Search AI leaks its chain-of-thought and internal formatting tags to users — ConsistentMap5775 · 2026-11-21
- SenseNova Loop 9B on a single RTX 4080: 256k context for 16GB VRAM GPUs — roofkid · 2026-11-21