1. OpenAI distillation defenses cut attack success to 0.2%
OpenAI distillation defenses have cut the success rate of attacks that try to pull protected reasoning out of GPT-6.2 Astra to 0.2%, the company said in research published on November 20, 2026. The same kind of attack had worked 54.0% of the time against the previous model, GPT-6 Astra, before the new safeguards were added.
The research centers on what OpenAI calls adversarial distillation, a practice in which outsiders try to copy the internal decision-making that makes a frontier model valuable. By sharply reducing how often that hidden reasoning can be extracted, OpenAI is trying to protect one of the main features that separates its most advanced systems from cheaper imitations.
The figures come from OpenAI's own testing and cover the company's newest models, GPT-6.2 Astra and GPT-6.2 Sol, though the headline results focus on Astra. They arrive as competition among AI developers increasingly turns on whether one lab can cheaply reproduce another's capabilities.
2. What is AI model distillation?
Distillation is the process of training one AI model on the outputs of another. A smaller or newer model, often called the student, learns by studying the answers produced by a larger and more capable model, the teacher. The technique is widely used across the industry and is a standard, legitimate way to build faster and cheaper models from more powerful ones.
The problem arises when the practice is used without permission. Adversarial distillation refers to systematically harvesting a rival's model outputs at scale in order to replicate its abilities, effectively copying years of research and billions of dollars in computing investment. OpenAI draws a clear line between sanctioned use of distillation and this kind of large-scale extraction, which it treats as a form of misuse.
Because modern AI models are offered through public chat products and developer interfaces, anyone with an account can send them prompts and collect responses. That openness is what makes them commercially useful, and it is also what makes them vulnerable to copying.
Distillation has become a central issue as the cost of building frontier models has climbed. Training a leading system requires vast amounts of computing power, specialized talent and carefully assembled data. A competitor that can approximate those results by studying another model's outputs can skip much of that expense, which is why developers have become increasingly protective of what their models reveal.
3. Why hidden reasoning is the prize
Advanced reasoning models do not simply produce an answer. Before responding, they work through a problem step by step in an internal chain of thought that users generally do not see in full. OpenAI refers to this as protected reasoning.
That hidden reasoning is especially valuable to anyone attempting distillation. A final answer tells a student model what the right output is. The reasoning behind it shows how to get there, revealing the strategies, intermediate steps and checks the teacher model used. Training on that richer material gives an attacker a far stronger learning signal than final answers alone could provide.
Protecting the reasoning therefore matters commercially. It is part of what distinguishes a frontier model from a lower-cost copy. If attackers could reliably extract it more than half the time, as the earlier figures for GPT-6 Astra suggest, the gap between an original model and its imitators could narrow quickly.
This is also why many AI companies show users only a summary of a model's thinking, or none at all. Keeping the full chain of thought private preserves the commercial edge while still allowing the model to use extended reasoning to reach better answers. The challenge is that the reasoning still exists inside the system, and determined attackers look for ways to bring it to the surface.
4. Two types of attack, two sets of results
OpenAI measured two kinds of extraction attacks. The first targets reasoning linked to an earlier request in a conversation, attempting to coax the model into revealing thinking it performed in a previous turn. Against GPT-6 Astra with its original safeguards, this attack succeeded 54.0% of the time. Against GPT-6.2 Astra with the updated defenses, the success rate fell to 0.2%.
The second type targets reasoning tied to the current request, trying to get the model to expose the thinking behind the answer it is producing at that moment. Here the success rate dropped from 22.9% on GPT-6 Astra to 0.9% on GPT-6.2 Astra.
Both results represent steep declines. The earlier-request attack, which had been the more effective of the two, is now the less effective one. The current-request attack, while far less successful than before, remains somewhat more likely to work than its counterpart.
5. How OpenAI hardened GPT-6.2 Astra
The company relied on two main defenses. The first is adversarial training. During training, the model is deliberately exposed to attempts to extract its reasoning so that it learns to recognize those attempts and resist them, while continuing to respond normally to legitimate requests.
The second defense operates at the level of OpenAI's application programming interface (API), the gateway developers use to access its models. Protected reasoning is now bound to the specific conversation history that produced it. Reasoning generated in one exchange is effectively locked to that exchange and cannot easily be carried into a different context and reused there.
That binding addresses a weakness that the earlier-request attack exploited. If reasoning from a previous turn cannot be transplanted into a new conversation designed to expose it, one of the main routes for extraction is closed off. The sharp drop in that attack's success rate, from 54.0% to 0.2%, reflects how central the technique had been.
The two defenses complement each other. Adversarial training makes the model itself more resistant to manipulation through cleverly worded prompts, while the API controls limit what an attacker can achieve even if a prompt gets past the model's own judgment. Layering protections in this way is a common principle in security engineering, since no single safeguard is likely to stop every attempt.
6. Enforcement and sharing threat data with rival labs
Technical safeguards are only one part of OpenAI's response. The company said it took weekly enforcement actions throughout 2026 against thousands of accounts tied to distillation campaigns, targeting the networks of users that attempt to harvest model outputs at scale.
OpenAI also shares indicators of compromise with other model providers. These are signals, such as patterns of behavior or account characteristics, that help identify the actors behind an attack. Passing them to other AI developers allows rival labs to spot the same groups if they turn their attention elsewhere.
That cooperation reflects the shared nature of the threat. Distillation is not a problem confined to a single company. An attacker blocked at one lab can simply move to another, so defenses that remain isolated within one provider leave the rest of the industry exposed. Sharing threat intelligence across companies that otherwise compete fiercely could encourage broader collaboration on AI security.
Account-level enforcement targets a different part of the problem. Distillation campaigns typically depend on spreading huge volumes of requests across many accounts to avoid detection and rate limits. Identifying and shutting down those networks raises the cost of running a campaign, even when individual prompts appear harmless on their own.
7. What the numbers mean for AI security
A 0.2% success rate does not mean the problem is solved. Attackers running large automated campaigns can send enormous numbers of requests, and even a small percentage of successful extractions can add up across millions of attempts. What the lower rate does accomplish is making each successful extraction far more expensive and time-consuming to obtain.
That shift in economics is the practical goal. Distillation is attractive because it promises frontier capabilities at a fraction of the cost of building them independently. If extraction becomes slow, costly and unreliable, the incentive to attempt it at scale weakens, and the advantage held by the original developer becomes more durable.
The results also highlight where pressure is likely to build next. The current-request attack rate of 0.9% is dramatically lower than the 22.9% recorded on GPT-6 Astra, but it remains above the 0.2% achieved for earlier-request attacks. Attackers looking for the weakest remaining point are likely to focus their efforts there.
The figures also illustrate how quickly the picture can change between model generations. A safeguard that seemed adequate for one release may prove weak once attackers study it closely, and the jump from 54.0% to 0.2% shows how much ground a determined defender can recover when it redesigns protections around the specific techniques that were working.
8. Why it matters for the AI industry
The research arrives shortly after the launch of GPT-6.2 Astra and GPT-6.2 Sol, OpenAI's newest models. As frontier systems become more capable, the value of their underlying reasoning grows, and so does the motivation for others to copy it.
For developers building on OpenAI's platform, stronger protections around reasoning reinforce confidence that the model they pay to use is not being quietly reproduced by competitors. For the wider industry, OpenAI's decision to publish extraction rates and share threat indicators sets a reference point that other labs may be asked to match, as customers and policymakers pay closer attention to how frontier capabilities are protected.
The contest between defenders and those attempting extraction is likely to continue. Each new safeguard invites new attack methods, and the remaining 0.9% gap on current-request reasoning shows that the work is not finished. For now, OpenAI's figures suggest the balance has tilted significantly toward the defenders.
