OpenAI says defences cut reasoning theft from 54% to 0.2% on GPT-6.2 Astra
OpenAI published 'Combatting adversarial distillation' on November 20, detailing its fight against the unauthorised use of a frontier model's outputs to train another model with weaker safeguards. In its own internal tests, the success rate for recovering protected reasoning fell from 54.0% to 0.2%.
📊 Scorecard: our Bullish call on $MSFT is scored from $484.52, the last close before it ran. Its first result comes after the next close.

The News
OpenAI published 'Combatting adversarial distillation' on November 20, detailing how it fights the unauthorised use of a frontier model's outputs to train another model with weaker safeguards — with particular emphasis on protecting the model's chain-of-thought, the intermediate reasoning steps that can carry information the final answer leaves out. Adversarial distillation is the practice of using one AI model's outputs to train another model, so that the second model gains the first model's abilities without its safety rules.
In internal adaptive black-box red-teaming — a test in which attackers try to break a system without seeing how it works inside — aggregated across six attacker models, the average success rate for recovering protected reasoning from an earlier request fell from 54.0% on GPT-6 Astra with its original safeguards to 0.2% on GPT-6.2 Astra with updated safeguards. For reasoning in the current request, the rate fell from 22.9% to 0.9%. With GPT-6.2 Astra and GPT-6.2 Sol, OpenAI says protected reasoning is tied to the conversation history that produced it — if an earlier message is rewritten, the API drops the reasoning that followed — and older model families can no longer read it.
The defences combine adversarial training, input and output classifiers, API controls and campaign-level monitoring. OpenAI also reports an internal study showing that a student model trained on exact reasoning traces generally learned more cyber capability than one trained on reasoning summaries — which it says shows that protecting traces can limit capability transfer. Weekly enforcement against distillation campaigns has ranged from thousands to tens of thousands of accounts, 'with the rate increasing significantly this past summer'. All of these figures are OpenAI's own internal evaluations, not independently verified. Microsoft (MSFT) is named here only as a read-through: OpenAI is private, and Microsoft is its largest commercial partner and investor.
From 54% to 0.2% — if the number holds, the reasoning hole is mostly closed.
Why It Matters
The numbers matter because they describe a real threat to the AI business. A frontier model's reasoning — the steps it takes to reach an answer — is one of the most valuable things it produces. If a rival can extract that reasoning and train a cheaper model on it, the rival gets the capability without paying for the research or the safety work. Cutting the extraction rate from 54% to 0.2% would close most of that hole.
The mechanism matters too. Tying protected reasoning to the conversation history that produced it — so that rewriting an earlier message drops the reasoning that followed — is a structural fix, not a filter that can be worked around. Combined with adversarial training and classifiers, it makes the attacker's job harder at every step, rather than relying on a single defence.
But the figures are OpenAI's own, and they have not been independently verified. The tests were run by OpenAI, on OpenAI's models, against attacker models OpenAI chose. The verdict is Bullish because the reported improvement is large and the mechanism is concrete — while remembering that an internal evaluation is a claim about a defence, not proof that the defence holds against every attacker.
Analyst's Take
Our call is Bullish. A fall from 54.0% to 0.2% in reasoning extraction, and from 22.9% to 0.9% in the current request, is a large reported improvement — and the conversation-history mechanism is the kind of structural change that is harder to work around than a filter.
What to watch next is whether attackers adapt. OpenAI's own enforcement numbers — thousands to tens of thousands of accounts a week, rising sharply over the summer — show the attacks are not stopping. The real test of these defences is not an internal red team, but the next wave of real distillation campaigns.
For investors, the read-through is Microsoft, ticker MSFT — OpenAI's largest commercial partner and investor. Protecting frontier reasoning protects the value of the models Microsoft sells through its own products, and a defence that holds up is worth more than a defence that leaks. But until an independent test confirms the figures, treat 0.2% as OpenAI's number, not a settled fact. Reader poll: Does this protect AI models — yes or no? Vote: Yes, big step, Attackers adapt, Too soon to tell.
Are you bullish or bearish on this story?
By The Numbers
FAQ
What is distillation?
Distillation is a technique for training a smaller AI model using the outputs of a larger one, so the smaller model learns the larger model's abilities. Adversarial distillation is doing that without permission, to build a rival model with weaker safety rules.
What is chain-of-thought?
Chain-of-thought is the intermediate reasoning an AI model works through before giving its final answer — the steps, not just the conclusion. Those steps can carry information the final answer leaves out, which is why OpenAI protects them.
What is red-teaming?
Red-teaming is testing a system by attacking it, to find its weaknesses before a real attacker does. OpenAI's test used six attacker models trying to recover protected reasoning without seeing how the system works inside.
What is an API?
An API is a connection that lets one piece of software use another's services. OpenAI's API is how developers use its models; the new defences include API controls that drop protected reasoning when a conversation is rewritten.
👁
Source: The Decoder