OpenAI says the success rate for extracting protected reasoning has been reduced from 54% to 0.2%
OpenAI published research on November 20 explaining its measures against adversarial distillation. In an internal adaptive red-team evaluation, the company tested whether attackers could recover protected reasoning (chain-of-thought) generated by the model when answering prior prompts. For attacks that extract protected reasoning from earlier requests, the average success rate fell from 54.0% for GPT-6 Astra with the original protections to 0.2% for GPT-6.2 Astra with updated protections.
The company said it has strengthened related protections in GPT-6.2 Astra and GPT-6.2 Sol, including binding protected reasoning to the conversation history that generated it and discarding later reasoning if prior messages are rewritten. The protections also include model behavior training, input and output classifiers, and API-level restrictions, and the company is investigating organized extraction activity.
Original source:
Title: Combatting adversarial distillation
Publisher: OpenAI
Publication date: 2026-11-20
URL: openai.com/index/combatting-adversarial-distillation
Source nature: company research announcement
Verification status: full text read in this instance, verification date 2026-11-20
Supporting content: supports the evaluation result from 54.0% to 0.2%, the model versions, the protection measures, and the statement that the extraction target was protected reasoning.
ABAB AI Insight
Protected reasoning refers to the model’s internally generated intermediate steps, which may contain information not disclosed in the final answer, and therefore has high value for training another model. OpenAI’s internal research showed that in cybersecurity tasks, student models trained with exact reasoning traces generally outperform student models trained only with reasoning summaries after supervised fine-tuning.
The company previously disclosed in September that it had blocked an organized effort to extract protected reasoning, with the core part attributed to individuals linked to Moonshot AI. This announcement aligns reasoning protections with standards for biological and cybersecurity safeguards, and mentions briefing Congress in February on DeepSeek-related activity.
The defenses are divided into the request level (model behavior, classifiers, API controls) and the activity level (account network investigations and enforcement). The company said it has taken weekly enforcement actions this year against related activity involving thousands to tens of thousands of accounts. The evaluation results come from internal adaptive black-box testing and are not an independent third-party audit.
Public information does not provide external verification data, specific attack samples, or a long-term effectiveness assessment. The company said it is applying some of the improvements to GPT-6 Astra.
ABAB News · Cognitive Law
The value of internal model reasoning traces for downstream model training is higher than using only final answers or summaries.
Binding reasoning to a specific conversation history can make cross-session extraction more difficult.
Defenses against adversarial distillation require combining request-level controls with investigations into organized activity.