🔍 Read the full analysis: Steps For Addressing Adversarial Distillation on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
OpenAI reports that updated safeguards in GPT-6.2 Astra reduced success in an internal test of protected-reasoning extraction from 54.0% to 0.2%. A separate company study found exact reasoning traces could transfer more cyber capability to student models than summaries, but OpenAI has not published full methods or numerical benchmark scores.
OpenAI says updated safeguards in GPT-6.2 Astra cut the success rate of an internal test for extracting protected reasoning from 54.0% to 0.2%, while a separate company study found that exact reasoning traces transferred more cyber capability to a student model than summaries in some tested conditions. The findings address adversarial distillation—using a model’s outputs to train another system—and are company-reported, with the extraction result limited to an internal evaluation whose full methods have not been released.
OpenAI says it compared GPT-6 Astra with its original safeguards against GPT-6.2 Astra with updated safeguards. In an internal adaptive evaluation using automated red-teaming, testers attempted to recover protected reasoning generated as the model answered an earlier prompt. The company reports that the test’s extraction success rate fell from 54.0% to 0.2%. These figures describe that evaluation; the source does not provide the sample size, detailed methodology or independent replication.
OpenAI also describes safeguards operating at more than one level: model training, classifiers and API protections address individual requests, while investigations look for patterns across requests and accounts that might indicate coordinated extraction. The company says investigation findings and automated red-team results feed into further training and testing. It did not provide details of specific investigations or enforcement actions in the supplied material.
In a separate cyber-capability study, OpenAI says researchers started each training run with the same pretrained student model and used supervised fine-tuning data drawn mainly from trajectories produced by Astra-level teachers on coding and defensive cybersecurity tasks. The data included task steps, tool interactions and final responses. Students trained on exact reasoning traces generally outperformed those trained on summaries after fine-tuning. Following additional lightweight reinforcement learning, the exact-trace student scored higher at the largest inference budget tested on both reported benchmarks, while the summary-trained student scored higher at lower budgets. OpenAI did not name the benchmarks or publish scores in the source material.
Combatting Adversarial Distillation
OpenAI reports stronger safeguards against attempts to extract protected reasoning, alongside evidence that exact reasoning traces can transfer cyber capability to student models.
How defenses work
OpenAI describes safeguards that look at individual requests and coordinated activity across requests and accounts.
Model training, classifiers and API-level protections aim to block reasoning extraction.
Patterns across requests and accounts can help identify organized extraction attempts.
Investigation findings and automated red-teaming feed back into training and testing.
What student models learn
OpenAI’s internal study compared training on exact reasoning traces with training on reasoning summaries.
- Students generally outperformed summary-trained students after supervised fine-tuning.
- After further reinforcement learning, the exact-trace student scored higher at the largest tested inference budget on both reported benchmarks.
- The summary-trained student scored higher at lower inference budgets after reinforcement learning.
- The study did not provide benchmark names or numerical scores.
Why it matters
Distillation can pass a teacher’s capabilities to a student model that may not have comparable safeguards.
Dual-use skills
Coding and defensive security work can also teach techniques applicable to attacks.
Sensitive research
Biological research can have both beneficial and harmful uses.
Why Reasoning Access Matters
The results point to a safety issue beyond copying a model’s final answers. If step-by-step reasoning transfers useful capability, training another system on those traces could give it skills that are not apparent from the final response alone. That matters if a student model is later deployed without protections comparable to the teacher’s safeguards. OpenAI’s study suggests summaries may transfer less capability in some tested conditions, but it does not establish that summaries prevent transfer or that the pattern applies across models.
The concern is relevant to dual-use fields such as cybersecurity and biology, where knowledge can support legitimate research or harmful activity. Coding and defensive security tasks, for example, can involve techniques that also have offensive applications. At the same time, restricting access to reasoning traces could affect legitimate research and model evaluation. The available results identify a potential risk and a possible mitigation, but do not quantify real-world extraction attempts or their frequency.
Products Worth Considering

Sony WH-1000XM6 Wireless Noise-Cancelling Headphones – Midnight Blue

Logitech MX Mechanical Mini for Mac Wireless Keyboard, Space Gray

Apple AirPods Pro (1st Generation) with MagSafe Charging Case
As an affiliate, we earn on qualifying purchases.
How Distillation Enters the Picture
In model distillation, a student model learns from data produced by a teacher. That data may include answers, judgments or reasoning. OpenAI uses “adversarial distillation” for unauthorized use of a frontier model’s outputs to train another system that may not have comparable safeguards. The issue is not limited to whether an output is copied: the training material could transfer capabilities to a separately deployed model.
For the cyber study, OpenAI compared students trained on exact teacher trajectories with students trained on reasoning summaries, then evaluated them after fine-tuning and after a further reinforcement-learning stage in cyber and coding environments. The report describes performance differences by training condition and inference budget, but does not provide benchmark names or numerical results. It also does not establish whether comparable effects occur with other teachers, student models, tasks or training recipes.
Limits of the Reported Tests
The supplied report does not state when the safeguards were introduced, whether they are available to all users, how extraction success was calculated or how many test attempts were conducted. No independent replication is described. The 0.2% figure should therefore be read as a result from OpenAI’s reported internal evaluation, not as a general estimate of the chance that protected reasoning can be extracted from models.
The cyber study also leaves key questions open. OpenAI does not give the benchmark names, numerical scores or full evaluation methods, and the reported comparison covers particular student models, tasks and inference budgets. The source does not show how findings generalize to other distillation methods or whether restricting exact traces changes real-world misuse. Details of investigations and enforcement are also absent.
Further Tests Across Models
OpenAI says it plans to continue testing capability transfer across different student models and distillation methods. The company says automated red-teaming and investigations will continue to inform safeguard development, but the source gives no schedule for additional results. Further reporting could clarify safeguard availability and describe how the company identifies and responds to coordinated extraction attempts.
For readers assessing the claims, useful next evidence would include a fuller account of the extraction test, its sample size and success criteria, alongside independent evaluations across different attack approaches. Published benchmark names and scores would also help show how much the exact-trace advantage varies across inference budgets and training setups. Until then, the reported findings remain promising but limited company evidence, rather than a broad demonstration that reasoning extraction has been prevented.
Key Questions
What is adversarial distillation?
It is the unauthorized use of a model’s outputs to train another model. OpenAI’s concern is that the resulting student may gain useful capabilities without safeguards comparable to those of the teacher.
What did OpenAI report about reasoning extraction?
OpenAI says an internal adaptive evaluation using automated red-teaming found that extraction success fell from 54.0% to 0.2% when it compared GPT-6 Astra’s original safeguards with GPT-6.2 Astra’s updated safeguards. The company has not provided the full methodology or independent replication in the supplied material.
Did exact reasoning traces improve student-model performance?
OpenAI says students trained on exact traces generally did better than summary-trained students after supervised fine-tuning. After additional reinforcement learning, the exact-trace student scored higher at the largest tested inference budget on both reported benchmarks, while the summary-trained student scored higher at lower budgets.
Are the findings independently verified?
The supplied source describes the results as OpenAI’s internal findings and does not report independent replication. It also omits the extraction test’s sample size and detailed methods, as well as benchmark names and scores for the cyber study.
Primary source: OpenAI · via ThorstenMeyerAI.com
