The transfer of AI capabilities raises the question of their safeguards

An OpenAI study examines the role of reasoning traces in distillation and describes its defenses against unauthorized extractions.

Data from programming tasks and defensive cybersecurity can convey skills useful for offensive uses. In a study on adversarial distillation, OpenAI examines this transfer and the particular role of exact reasoning traces. The lab also describes the safeguards deployed to limit their extraction, from individual exchanges to networks of coordinated accounts.

Distillation is a common technique: a model learns from the answers, judgments, or reasoning produced by another. The campaigns targeted here are those the provider considers unauthorized. The risk described goes beyond intellectual property: certain capabilities could be reproduced in a model lacking protections comparable to those that govern the original model. What the intermediate steps change

The experiment starts from the same pre-trained model to compare two supervised training runs. The first uses the elements normally visible to users, including final answers, reasoning summaries, and interactions with tools. The second provides access to exact traces: the intermediate steps generated by the teacher model during problem solving. The data come mainly from code and defensive cybersecurity tasks carried out by models at the Astra level.

After this training, the student that received the exact traces generally performs better. The comparison shifts after an additional reinforcement learning phase: on ExploitBench and SEC-bench Pro, it remains ahead at the largest compute budget tested, while the student trained on summaries does better at lower budgets. The benefit of full traces therefore also depends on the execution conditions.

A paper published by Anthropic had already compared these two types of data on SWE-bench Verified. The model trained with the full reasoning achieved 80.4%, versus 71.8% with their summaries. These separate experiments study the value of traces as training material, without establishing that distillation would reproduce all of the teacher model’s capabilities. Defenses tested against adaptive attacks

The defense described combines training intended to refuse disclosure of protected reasoning, input and output filters, and restrictions in the API. The controls also apply to the arguments sent to tools: a trace could appear there without appearing in the final response. The protections cover both streaming responses and those returned all at once.

For GPT-6.2 Astra and GPT-6.2 Sol, the protected reasoning is tied to the history that produced it. If a previous message is rewritten, the API discards the traces that follow it. The provider also says it prevents older model families, which are less protected, from reading the protected reasoning of these two versions.

In an internal adaptive black-box evaluation, six attacking models adjust their attempts according to the responses obtained. To retrieve the reasoning for a previous request, the average success rate rises from 54% on GPT-6 Astra with its original protections to 0.2% on GPT-6.2 Astra with the new defenses. For the current request, it goes from 22.9% to 0.9%. These figures measure extraction in this protocol, not the overall success rate of the distillation campaigns. The comparison changes both the model version and the protections surrounding it. The network of accounts as the unit of investigation

At the scale of campaigns, investigations cross-reference request signals and account signals to detect coordinated activity. The lab reports weekly interventions involving thousands to tens of thousands of accounts, with a marked increase over the summer. The attacks discovered then feed defensive training and regression testing. Indicators are also shared with other providers to look for related activity.

The publication also links these defenses to an explicit geopolitical stance. The group supports U.S. measures targeting foreign actors who illicitly extract model capabilities, as well as stronger controls on hardware exports. It presents protection against distillation and restriction of access to compute as two complementary means of preserving U.S. leadership in AI.