Skip to content
OpenAI News

OpenAI details defenses against adversarial distillation and protected reasoning extraction

Combatting adversarial distillation

OpenAI published its defenses against adversarial knowledge distillation, tightening protection for protected reasoning in GPT-6.2 Astra and GPT-6.2 Sol. In internal adaptive black-box tests, the attack success rate for extracting reasoning from earlier requests fell from 54.0% on GPT-6 Astra to 0.2% on GPT-6.2 Astra, and for extracting reasoning from the current request from 22.9% to 0.9%.

Why it matters: OpenAI extended extraction defenses to tool-call arguments and cross-session reuse, and used samples from coordinated attack investigations to improve training and regression tests.

Read the original ↗Export Markdown