OpenAI: reasoning extraction from GPT-6.2 Astra succeeds in 0.2% of attacks, versus 54.0% for GPT-6
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
Rewriting an earlier message in the API discards every reasoning step generated after it, so a chain can no longer be assembled piecemeal from the conversation history.
On November 20 OpenAI disclosed a multi-layered defense against adversarial distillation — the coordinated siphoning of a model's outputs and reasoning chains in order to train a student model. According to OpenAI, a July campaign involved about 16,000 reasoning-extraction requests from more than 4,000 users. The main group, the company says, was in contact with people linked to Moonshot AI. In GPT-6.2 Astra and GPT-6.2 Sol, protected reasoning is tied to the conversation history. When a message is rewritten, the API discards all subsequent reasoning, and models from older families cannot read such chains. In an internal red team test, the success rate of extracting reasoning from a previous request fell from 54.0% for GPT-6 Astra to 0.2% for GPT-6.2 Astra. The protections include input and output classifiers, validation of tool-call arguments and account-level campaign analysis.
Sources 2