Dek: OpenAI published the before-and-after numbers for its latest safeguards and they look like a fix. The same post explains why the fix has to be repeated every week, by everyone.
The OpenAI numbers: 54% down to 0.2%
Start with the pair that carries the announcement. Against GPT-6 Astra, an attacker trying to recover protected reasoning from an earlier request succeeded 54.0% of the time. Against GPT-6.2 Astra with the updated safeguards, the same attack succeeded 0.2% of the time.
Extraction from the current request fell from 22.9% to 0.9%.
Those figures come from OpenAI's internal adaptive black-box evaluation, aggregated across six attacker models the company tested against. That's a vendor testing its own defences with its own red team, not an independent audit, and the company says both rates landed below 1% per attack type with the new safeguards in place.
OpenAI also says selected pieces of the new protection are going back to GPT-6 Astra. How much of that gap survives outside OpenAI's lab?
What adversarial distillation means in plain English
Distillation is ordinary machine learning. A small model learns by studying a bigger model's answers. It's how most teams build a cheap model that behaves like an expensive one.
Adversarial distillation is the same idea without permission. Someone collects a frontier model's outputs, trains a student on them, and deploys the result without the safeguards that came with the teacher. The capability travels. The safeguards stay home.
That framing matters for how you read the rest of the announcement. The company isn't arguing that distillation is theft in general. It's arguing that a specific use of it, at scale, against a model that costs billions to train, is a problem it now has to police.
Why chain-of-thought traces are the valuable part
A finished answer is the least interesting thing to steal. The chain-of-thought is where the work happens: intermediate steps, self-corrections, the branch a model abandons before settling.
OpenAI calls this "protected reasoning" and says the traces contain information the final answer leaves out. For anyone building a student model, that difference is the whole product.
The company ran an internal cyber-capability transfer study to test the claim. A student fine-tuned on exact reasoning traces generally beat one trained on summaries after supervised fine-tuning, and still scored higher at the largest inference budget after a light reinforcement learning round. The summary-trained student led at lower inference budgets, which is the one part of the result that cuts the other way.
So reasoning extraction pays off when the attacker can afford the compute. That's the case OpenAI is defending against.
The four OpenAI defences, in the order they act
Each layer catches what the one before it misses, which is why the company lists them as a sequence rather than a menu.
First, adversarial training teaches the model to recognise extraction attempts and refuse them. That's the layer a user feels, because a refusal is visible.
Second, input classifiers look at what arrives and flag prompts that smell like data collection. Third, output classifiers cover assistant messages and tool-call arguments, in streaming and non-streaming responses both. Tool-call arguments are the easy thing to forget, since they look like plumbing.
Fourth, API protections tie a model's protected reasoning to the conversation history that produced it. Rewrite an earlier message and that reasoning vanishes. Older model families with weaker safeguards also can't read GPT-6.2 Astra and Sol reasoning at all.
How OpenAI hits adversarial distillation campaigns
Single requests are the small end of this. OpenAI says it investigates the accounts behind coordinated extraction attempts, enforces against the network rather than the individual account, and shares indicators of compromise with other providers.
The volume is the number worth remembering. The company bans anywhere from thousands to tens of thousands of accounts weekly, and says the rate has risen sharply since the summer.
That's an enforcement posture, not a product feature. It also explains why OpenAI keeps publishing these posts: the argument only works if other labs adopt the same practices, and adoption is easier when the alternative is a weekly fight.
The national security case OpenAI is making
OpenAI links the technical work to policy, and the links are specific.
It briefed Congress in February on activity indicating DeepSeek was trying to distill US frontier models. In September it reported disrupting a campaign to extract protected reasoning, attributing a core part of it to individuals associated with Moonshot AI, the developer of Kimi. That's an attribution to people, not a charge against the company.
The company says it supports the Chip Security Act, the AI OVERWATCH Act and the MATCH Act, along with legislation giving the US government more tools to identify and sanction foreign actors that extract capabilities. Those are bills it backs, not law.
For a reader outside the US, the practical consequence is simpler. Export rules govern hardware. This argument extends the same logic to behaviour, which is harder to inspect at a border.
What the OpenAI numbers don't prove
A 0.2% attack success rate is a measurement of attacks OpenAI designed, run against models OpenAI built, scored by OpenAI. Nobody outside has reproduced it.
The number also isn't a theft rate. It says how often a particular class of attack worked in a lab setting over a fixed set of attempts. Real campaigns adapt, and the post itself describes an ongoing fight rather than a closed case.
Anthropic is running a parallel effort, according to The Decoder, banning tens of thousands of accounts and reportedly succeeding in 90% to 99% of cases, and has called for a coordinated industry and policy response. Two labs reporting similar numbers isn't verification. It's two vendors agreeing that the problem is real.
Better, not solved.
OpenAI's earlier September disclosure described attackers moving encrypted reasoning between conversations and asking the model to reveal it in readable form, which is a reminder that each fix reshapes the attack rather than ending it.
What OpenAI wants other developers to copy
The closing argument of the post is a checklist aimed at other model developers, and it's worth reading as the actual point of publication.
Protect private reasoning across the whole interaction, tool-call arguments included. Keep logs of successful attacks and turn them into regression tests, so a fixed hole stays fixed. Pair per-request defences with campaign-level enforcement, since a classifier can't see a network of accounts working in parallel. Coordinate safeguard requirements with deployment partners.
Nobody sells any of that as a product. It's a set of practices, published because the company's own numbers only mean something if its competitors are measuring the same thing.






