OpenAI official logo (public domain, Wikimedia Commons) — CryptoBriefing brand treatment
The company's distillation research shows sharp drops in attack success rates between GPT-6 Astra and GPT-6.2 Astra
OpenAI says its updated safeguards have pushed the success rate of attacks that pull protected reasoning out of GPT-6.2 Astra down to 0.2%. In the previous model, GPT-6 Astra, the same category of attack worked 54.0% of the time.
The figures come from research OpenAI published on November 20, 2026, focused on what it calls adversarial distillation. The goal is to stop outsiders from quietly siphoning off the internal decision-making that makes its models valuable.
What OpenAI is defending against #
Distillation, in AI terms, means training one model on the outputs of another. Done with permission, it is a standard way to build smaller, cheaper systems. Done adversarially, it starts to look a lot like copying someone’s homework at industrial scale.
The specific target here is protected reasoning. That is the step-by-step thinking a model works through before giving an answer, which OpenAI keeps hidden from users. If an attacker can coax that hidden chain of thought out of the model, they get a training signal that is far richer than the final answer alone.
OpenAI’s research measured two kinds of extraction attempts. For reasoning tied to earlier requests in a conversation, attack success fell from 54.0% on GPT-6 Astra to 0.2% on GPT-6.2 Astra. For reasoning tied to the current request, success dropped from 22.9% to 0.9%.
How the safeguards work #
According to the research, OpenAI is leaning on two main defenses. The first is adversarial training, which means deliberately exposing the model to extraction attempts during training so it learns to resist them.
The second defense lives at the API level. OpenAI ties a model’s reasoning to specific conversation histories. Put simply, the reasoning generated in one exchange is locked to that exchange, so an attacker cannot easily pick it up and replay it in a different context to fish out more.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
OpenAI says it has taken weekly enforcement actions throughout 2026 against thousands of accounts involved in distillation campaigns. The company also says it shares indicators of compromise with other model providers to let rival labs spot the same actors on their own platforms.
The release context #
The research landed shortly after OpenAI released GPT-6.2 Astra and GPT-6.2 Sol, which came out around November 17, 2026. The research covers both models, though the headline extraction figures focus on Astra.
Alongside the launch, OpenAI cut Astra’s API pricing by 20%.
What this means #
For OpenAI, the stakes are straightforward. Hidden reasoning is part of what separates a frontier model from a cheaper imitation. If competitors or bad actors can extract it at a 54.0% clip, the gap between the original and the copy shrinks fast, and so does the payoff from building the original. Getting that number down to 0.2% on the earlier-request attack does not eliminate the problem. At scale, even a small success rate can add up for a determined attacker running many attempts. But it changes the economics, making each successful extraction far more expensive to obtain.
The research also suggests these defenses may matter beyond competitive protection. The deep research findings note that lower adversarial risk could boost confidence in deploying OpenAI’s systems in sensitive areas such as cybersecurity and biological research. In those fields, leaking a model’s internal reasoning is not just a business concern. It could hand capabilities to people who should not have them.
The information-sharing piece may end up being the most consequential part. Distillation is an industry-wide threat, and an attacker blocked at one lab can simply move to another. By distributing indicators of compromise to other model providers, OpenAI is treating this less like a private fight and more like a shared security problem. The deep research findings suggest this could foster a more collaborative environment that strengthens security across the sector.
The current-request success rate of 0.9%, while far below the earlier 22.9%, is also still higher than the earlier-request figure, which hints at where pressure may concentrate next.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.