OpenAI published a guide to combatting adversarial distillation on November 20, 2026, detailing stronger protections for the internal reasoning generated by GPT-6.2 Astra and GPT-6.2 Sol. The approach combines model training, classifiers, tool-call checks, API restrictions and investigations into coordinated extraction campaigns.
OpenAI reports that attacks targeting reasoning from an earlier request succeeded 0.2% of the time against GPT-6.2 Astra with updated safeguards, compared with 54.0% against GPT-6 Astra with its original protections. Attacks targeting reasoning from the current request fell from 22.9% to 0.9%.
The useful contribution is the implementation detail behind those figures. OpenAI describes protections at several points where reasoning could leave the system, including tool calls and streamed responses. That makes this a safety architecture update, with practical lessons beyond the headline percentages.
What OpenAI’s Extraction Results Measure
OpenAI’s evaluation used adaptive black-box red teaming, with results aggregated across six attacker models. In a black-box evaluation, attackers probe the target through its exposed interface; adaptive testing lets them adjust their attempts in response to what works.
| Extraction target | GPT-6 Astra, original safeguards | GPT-6.2 Astra, updated safeguards |
|---|---|---|
| Protected reasoning from an earlier request | 54.0% | 0.2% |
| Protected reasoning from the current request | 22.9% | 0.9% |
Average attack success rates reported by OpenAI. Lower is better. Source: OpenAI.
These are vendor-internal results, not independently reproduced measurements. Neither the announcement nor The Decoder’s accompanying coverage supplies independent validation.
The GPT-6.2 System Card, published November 17, had already described stronger unauthorized-distillation protections. It reports extraction success below 1% for both earlier-turn and current-turn reasoning on both GPT-6.2 models. That is consistent with the guide’s Astra figures, but it remains another OpenAI document, not a separate evaluator’s confirmation.
The comparison also changes two things at once: the model version and its safeguards. It measures the combined move from GPT-6 Astra with original protections to GPT-6.2 Astra with updated protections. It cannot tell a reader how much of the improvement came from training, classifiers, API controls or another component.
I’d read the table as evidence that OpenAI’s updated package performed substantially better in its chosen evaluation, not as a measurement of any single defense.
Why Exact Reasoning Traces Matter
Protected reasoning is the intermediate chain of thought a model generates while working toward an answer. It is distinct from the final response or a reasoning summary shown to a user.
Distillation trains a student model using material produced by a teacher model. OpenAI’s concern is that an actor could use a frontier model’s outputs to acquire capabilities without carrying over the safeguards that constrain the original service. The company places particular emphasis on exact reasoning traces because they can contain information omitted from visible answers.
OpenAI supports that concern with an internal cybersecurity distillation study. Researchers started each training run from the same pretrained student model and used trajectories from Astra-level teachers working on coding and defensive cybersecurity tasks. They compared training with exact reasoning traces against training using the normally visible material, including reasoning summaries.
The results require some care:
- After supervised fine-tuning, the exact-trace student generally outperformed the summary-trained student.
- After additional lightweight reinforcement learning, the exact-trace student still scored higher at the largest inference budget tested on both benchmarks shown.
- At lower inference budgets, the summary-trained student scored higher.
Those conditions matter. The study supports OpenAI’s argument that exact traces can transfer additional capability under some training and inference settings. It does not show that exact traces always produce a stronger student.
The consequential point is that the teacher’s original tasks were coding and defensive cybersecurity work. OpenAI argues that skills learned across many permitted interactions can generalize to riskier uses. Protecting reasoning therefore addresses a different problem from simply refusing an explicitly malicious request.
Request-Level Defenses Cover More Than Chat Text
OpenAI separates its safeguards into request-level protections and campaign-level enforcement. At the request level, the design places checks around several possible routes for exposing protected reasoning.
Training and Classifiers
Targeted model training aims to make the model resist requests to reveal protected reasoning. Input classifiers examine incoming material for extraction attempts, while output classifiers check generated material for possible disclosures.
These layers serve different purposes. Training shapes the model’s behavior before a separate detector intervenes. Classifiers provide additional checks around that behavior. OpenAI presents the combination as mutually reinforcing, with attacks discovered through investigations feeding back into training and testing.
Tool-Call Inspection
OpenAI’s classifier coverage includes tool-call arguments, not only messages displayed to users.
That detail deserves attention from anyone building an agent. A model can send information to a tool as part of completing a task, so a final-answer filter alone leaves another output path to consider. Checking tool arguments extends the protection to content passed into those interactions.
My view is that this is one of the guide’s most useful design lessons: a tool call is an output boundary, even when it never appears as ordinary chat text. The same architectural reasoning applies to other sensitive information, although OpenAI’s announcement focuses on protected reasoning.
API Controls and Encrypted Reasoning
OpenAI also describes API-level protections for protected reasoning, including binding it to conversation history and preventing older models from reading it.
The practical distinction is between carrying protected state through an authorized conversation and treating that state as freely reusable text. These controls address how reasoning is handled across requests and model access, a boundary that a refusal instruction alone cannot enforce.
The system card separately identifies protections for encrypted reasoning. Encryption and permitted reuse are different concerns: protecting a representation does not remove the need to control where and how the service accepts it.
Holding Streamed Responses
The system card says additional checks can hold streamed responses when they might expose private model reasoning.
That creates a straightforward product tradeoff. Holding output gives a safeguard an opportunity to intervene before potentially sensitive material reaches the recipient. It can also delay when the application receives visible content. OpenAI provides no latency measurement for this behavior in the supplied evidence, so there is no supported number to attach to that tradeoff.
Frequently Asked Questions
3 questions
1What Is Protected Reasoning?
It is the intermediate chain of thought generated internally while a model works toward an answer. It is different from the final response or a user-visible reasoning summary.