OpenAI has tightened protections against unauthorized distillation in which an actor trains a second model on the outputs of an advanced model and deploys it with weaker safeguards. The company says this can transfer capabilities that the original model would use only under restrictions.
The risk is particularly significant in dual-use fields. Skills used to find software vulnerabilities can also be useful in attacks, while capabilities that assist biological research could be applied to developing a harmful pathogen. OpenAI gives the example of a distilled model helping with a cyberattack that the original model or its safeguards would have refused to support.
Protecting internal reasoning
Distillation is itself a common machine-learning technique: one model learns from another model’s answers, reasoning or evaluations. OpenAI says unauthorized use raises safety concerns in addition to intellectual property issues because the student model may acquire capabilities without inheriting the original model’s safeguards.
The company is particularly focused on preventing the disclosure of protected internal reasoning, known as chain-of-thought. These intermediate steps, generated while a model works toward an answer, may contain information omitted from the final response and can therefore be valuable training data.
OpenAI says it strengthened these protections in GPT‑6.2 Astra and GPT‑6.2 Sol. In an internal automated red-team evaluation, it tested whether attackers could recover reasoning generated while answering an earlier request. The average attack success rate fell from 54.0% against GPT‑6 Astra with its original safeguards to 0.2% against GPT‑6.2 Astra with updated protections.
What the cybersecurity study found
OpenAI conducted an internal study to assess how much cybersecurity capability could be transferred through exact reasoning traces. Teacher models worked on coding and defensive cybersecurity tasks; the baseline for comparing student models used material normally visible to users, including final responses, reasoning summaries and tool interactions.
Each training run began with the same pretrained student model. The supervised fine-tuning data consisted primarily of complete task trajectories from Astra-level teacher models, including tool interactions and final responses. The students were evaluated after fine-tuning and again after a lightweight round of reinforcement learning in coding and cybersecurity environments.
After fine-tuning, the student trained on exact reasoning traces generally outperformed the one trained on summaries. After reinforcement learning, the exact-trace student still scored higher on both benchmarks shown at the largest inference budget tested, while the summary-trained student performed better at lower budgets.
The company says the results suggest that protecting reasoning traces can limit the transfer of cyber capabilities even when the original task is permitted. OpenAI plans to study the effect further across different student models and training methods. It also noted that Anthropic previously found a similar advantage for full reasoning transcripts on SWE-bench Verified, a software engineering benchmark.
Layered defenses and testing
The company says it addresses both individual requests and coordinated campaigns. It trains models on adversarial examples to recognize attempts to extract protected reasoning while preserving answers to legitimate requests for explanations. Training examples come from internal red-teaming, investigations and responsible disclosures by outside researchers or industry partners.
Input classifiers look for extraction attempts, while output classifiers check assistant messages and tool-call arguments for leaks in both streaming and non-streaming responses. Checking tool calls matters because a model can disclose reasoning there even if its final answer does not. Suspected leaks can be blocked before reaching the caller.
API-level restrictions govern how protected reasoning can be reused across conversations and models. OpenAI’s September disclosure described attacks that moved encrypted reasoning between conversations and asked a model to reveal it in readable form. The company says it strengthened protections across user, workspace, organization and model-family boundaries and closed a pathway that could expose another user’s encrypted reasoning.
For GPT‑6.2 Astra and GPT‑6.2 Sol, protected reasoning is tied to the conversation history that produced it. The model can build on earlier reasoning as messages are added, but if an earlier message is rewritten, the API drops the reasoning generated after it. Older model families with weaker safeguards are also prevented from reading the two newer models’ protected reasoning.
To test its defenses, OpenAI replays previously successful attacks and uses automated attackers that adapt their methods to the system’s responses. Some testers receive only the information available to an external user, while others receive additional system information.
In adaptive black-box testing, the average success rate against GPT‑6.2 Astra’s updated safeguards was below 1% for attempts to extract reasoning associated with either an earlier request or the current one. OpenAI says both rates were substantially lower than for GPT‑6 Astra with its original safeguards; it is also applying selected improvements to that model.
Campaigns and policy position
OpenAI says it has detected multiple unauthorized distillation campaigns on its platform during the year. Its weekly enforcement actions have involved thousands to tens of thousands of accounts, with the rate rising significantly over the past summer. The company also tracks the prevalence of abuse and the time between detection and enforcement, aiming to reduce the volume and value of data obtained and increase the cost of attacks.
It uses findings from investigations to improve defenses and shares indicators of compromise with other model providers. With deployment partners, it applies shared safeguard and enforcement requirements to services offering OpenAI models.
The company also identifies a national security risk when foreign actors reproduce capabilities from U.S. models. It says it briefed the U.S. Congress in February on activity indicating that DeepSeek was attempting to distill American frontier models. In September, it reported disrupting a campaign to extract protected reasoning and attributed a core part of the activity to individuals associated with Moonshot AI, the developer of Kimi.
OpenAI supports giving the U.S. government more tools to identify and sanction foreign actors and says it has told Congress it supports the Chip Security Act, the AI OVERWATCH Act and the MATCH Act. In its view, export controls and anti-distillation defenses are complementary: the former can restrict access to hardware needed to train advanced models, while the latter make it harder to acquire capabilities from models already trained.
Recommendations for other developers
OpenAI recommends protecting private reasoning throughout an interaction, including in tool calls and the reuse of encrypted data. It says logs of successful attacks should be retained for regression testing and, consistent with privacy rules, used to improve training and classifiers. The company also recommends adaptive testing, combining request-level protections with campaign investigations, and coordinating enforcement with partners.
OpenAI says it has previously shared findings through the Frontier Model Forum and appropriate government information-sharing channels without making details that attackers could exploit public.



