OpenAI says it is fighting a growing problem it calls adversarial distillation: people using its models’ answers, and especially their hidden reasoning, without permission to train rival AI systems. In a post published on 20 November 2026, the company set out how it detects these campaigns, what it has changed in its new GPT-6.2 Astra and GPT-6.2 Sol models, and why it treats the issue as a safety and national security risk as well as a commercial one.
Here is what the term means, why the reasoning matters so much, and what we can and cannot verify.
What is distillation, and when does it become “adversarial”?
Distillation is an ordinary machine-learning technique. A smaller or newer model, the “student”, is trained on material produced by a stronger one, the “teacher”. That material might be the teacher’s final answers, its step-by-step working, or its judgements about which answer is better. Labs use it on their own models all the time to make cheaper, faster versions.
OpenAI uses “adversarial” for when it is done without authorisation, usually at scale, by someone who wants the teacher’s abilities without paying for or building them. Its argument goes beyond intellectual property. A student model can pick up a frontier model’s skills without the safety rules that come with them. The examples OpenAI gives are cybersecurity and biology, where the same knowledge can help defenders or attackers. A copied model might then help with requests the original would have refused or blocked.
Why is OpenAI so focused on the reasoning?
Modern “reasoning” models work through intermediate steps before they answer, often called the chain of thought. OpenAI calls this the model’s protected reasoning. Users normally see a summary of it, not the exact trace. OpenAI says those raw traces often contain things the final answer leaves out, which makes them especially valuable as training data for someone else’s model.
To test this, OpenAI says it trained the same starting student model on trajectories from Astra-level teacher models doing coding and defensive cybersecurity tasks. One version learned from what users can normally see (final responses, reasoning summaries and tool use). The other also learned from the exact reasoning traces. After fine-tuning, the student given exact traces generally did better. After a further light round of reinforcement learning, it still scored higher at the largest compute budget tested, although the summary-trained student did better at lower budgets. OpenAI’s conclusion is that hiding exact traces can limit how much cyber capability leaks across, even when the original tasks were allowed. This is the company’s own internal research and has not been independently replicated.
What has OpenAI changed in GPT-6.2?
The defences work at two levels. For individual requests, OpenAI says it has:
- Trained the models to resist extraction. Adversarial training teaches them to spot attempts to pull out protected reasoning and to refuse, while still giving normal explanations when people ask for them. The training examples come from attacks found by OpenAI’s automated red-teaming, its investigations, and disclosures from outside researchers.
- Added classifiers on the way in and out. Input classifiers look for extraction attempts. Output classifiers look for leaked reasoning in replies and in tool-call arguments, because a model could reveal its reasoning in a tool call without putting it in the visible answer.
- Tightened the API. In GPT-6.2 Astra and Sol, protected reasoning is tied to the conversation that produced it. If an earlier message is rewritten, the API drops the reasoning that followed it. Older model families with weaker safeguards can no longer read the new models’ reasoning. OpenAI also says it closed a pathway that could have exposed another user’s encrypted reasoning, after attacks it described in September moved encrypted reasoning between conversations.
On top of that, OpenAI says it investigates coordinated campaigns, acts against whole networks of accounts, and shares indicators of compromise with other model providers. It says its weekly enforcement actions have ranged from thousands to tens of thousands of accounts, and that the rate rose significantly over the summer.
The headline figures come from OpenAI’s own adaptive “black-box” red-teaming, where automated attackers see only what an outside user would see. Across six attacker models, the average success rate for extracting reasoning from an earlier request fell from 54.0% against GPT-6 Astra with its original safeguards to 0.2% against GPT-6.2 Astra. For reasoning from the current request, it fell from 22.9% to 0.9%. OpenAI says it is now applying some of the improvements to GPT-6 Astra too.
Who does OpenAI say is doing it?
OpenAI frames this partly as a US-China issue, and these are its claims, not established findings. It says that in February it briefed Congress on activity indicating that DeepSeek was attempting to distil US frontier models. It also says that in September it disrupted a campaign to extract its models’ protected reasoning and attributed a core part of it to individuals associated with Moonshot AI, the developer of Kimi. The sources we read did not include a response from either company, and we have not seen independent evidence for the attributions. OpenAI says it plans to share more about recent campaigns.
The company also uses the post to back US legislation, including the Chip Security Act, the AI OVERWATCH Act and the MATCH Act. It argues that export controls limit access to the hardware needed to train frontier models, and distillation defences make it harder to copy capabilities from models that already exist.
What does this mean for developers and everyday users?
For most people using ChatGPT or the API normally, OpenAI says nothing should change. Legitimate requests for explanations are meant to keep working, and reasoning summaries are still there. Developers building on GPT-6.2 should note that editing an earlier turn in a conversation now throws away the reasoning that came after it, and that reasoning can’t be carried across to older model families.
Some things are still unknown. All the effectiveness numbers come from OpenAI’s internal tests, so outsiders can’t yet check how well the defences hold up against determined attackers in real use. OpenAI hasn’t published false-positive rates, so we don’t know how often genuine users might be flagged as extractors. And there is a tension that will keep coming up: AI companies trained their own models on huge amounts of other people’s work, yet now want strict limits on how their outputs are reused. Expect that argument to grow as more companies release reasoning and agent-style models whose working is worth copying.
