parallelquant
November 20, 2026 · The Decoder

OpenAI cuts reasoning-extraction attacks on GPT-6.2 from 54% to 0.2%

OpenAI says new safeguards make it far harder to extract the hidden reasoning of GPT-6.2 Astra and Sol; in adaptive red-team tests, extraction success dropped from 54% to 0.2%. The company also says it suspends thousands to tens of thousands of accounts each week over coordinated distillation campaigns.

Why it matters: Hidden chains of thought are now treated as core intellectual property, because rivals can distill them to train cheaper models. The account-suspension numbers show distillation has become an industrial-scale threat, especially as GPT-6.2's cheaper API pricing makes bulk querying more economical. Expect other labs to follow with similar defenses, which also reduces how much outside researchers can inspect model reasoning.

Related updates