OpenAI protects GPT-6.2 Astra's reasoning: theft success rate falls from 54% to 0.2%
If you run a strong model through an API and train your own on its answers, the copy gets the skills without the restrictions, such as help with cyberattacks. OpenAI is closing that route for chains of thought, the hidden steps a model takes before it answers. In a test with automated attackers, the success rate for extracting reasoning from a previous request fell from 54.0% for GPT-6 Astra to 0.2% for GPT-6.2 Astra. If you serve a model through an API, check for leaks in tool-call arguments too.
AI-processed from OpenAI Blog; edited by Hamidun News
Imagine spending years building an expensive model with strict restrictions: it helps defenders find vulnerabilities in code but refuses to help with real attacks. Then someone runs it through the API at scale, collects the answers and trains their own model on them. The skills are copied; the restrictions are not.
OpenAI calls this adversarial distillation, and on 20 November it described how it is fighting it. The headline number: in a test where automated AI attackers tried to extract a model's hidden reasoning from a previous request, the share of successful attacks fell from 54.0% for GPT-6 Astra to 0.
2% for GPT-6.2 Astra. The source calls this share ASR, or attack success rate: the lower it is, the better the defense.
If you expose a model externally through an API, this is a call to action: protect reasoning the same way you protect final answers.
Why steal "thoughts" instead of answers
Distillation is an ordinary technique in itself: one model (the "student") learns from the data of another (the "teacher"), for example from its answers, its reasoning, or its ratings of which answers are better. Problems begin when this is done without permission. Beyond intellectual property concerns, the student picks up the capabilities but does not inherit the safeguards. This is especially dangerous in dual-use areas such as cybersecurity and biology: a model that helps find vulnerabilities can also help exploit them.
Chains of thought are of particular interest to attackers: these are the intermediate steps a model generates internally before it answers. They can contain information that is absent from the final text, which makes them valuable training data.
OpenAI tested this on cyber skills. The same pretrained student was fine-tuned on trajectories from Astra-level models: programming and defensive cybersecurity tasks, with all steps, tool calls and final answers. The baseline was what a user normally sees: answers, brief summaries of the reasoning, and tool use. It was compared with a student that received the exact chains of thought.
After ordinary fine-tuning, the student trained on exact chains was stronger overall than the student trained on summaries. After a light round of reinforcement learning in cyber and coding environments, it still won at the largest inference budget on both tests shown, while the summary-trained student led at smaller budgets. The company's conclusion: protecting exact chains can limit the transfer of cyber skills, even when the underlying tasks are permitted by policy. OpenAI is still studying how many skills transfer to other students and under other training recipes.
How the defense works
The defense works on two levels. The first is the request level:
- the model is fine-tuned on real successful attacks from internal automated red-teaming, investigations and reports from external researchers: it refuses to reveal protected reasoning but still explains its solutions normally when the request is legitimate;
- input classifiers catch attempts to extract reasoning, while output classifiers look for leaks in assistant messages and in tool-call arguments, in both streaming and non-streaming responses. Arguments matter because a model can "dump" its reasoning there while the final answer contains none of it;
- in the API, protected reasoning is tied to the conversation history. If an earlier message is rewritten, the API discards the reasoning that came after it. Older model families with weak protection cannot read the reasoning of GPT-6.2 Astra and Sol. The boundaries between users, workspaces, organizations and model families have also been tightened, and a path that could have exposed other people's encrypted reasoning has been closed.
The second level is the campaign level. OpenAI looks for patterns in request and account signals that point to coordinated extraction, and blocks entire networks of accounts. Each case it investigates becomes an example for training and testing the defenses, and the company shares indicators of compromise with other providers. According to OpenAI, weekly blocks cover from thousands to tens of thousands of accounts, and the pace picked up noticeably this summer.
What the measurements showed
The check ran as an adaptive black-box test: attackers see the system the way an outside user does and adjust their techniques to its responses. This finds weaknesses that static tests, with a fixed set of old attacks, miss. The result is averaged across six attacker models. For extracting reasoning from a previous request, success fell from 54.0% for GPT-6 Astra to 0.2% for GPT-6.2 Astra. For the current request, it fell from 22.9% to 0.9%. OpenAI is now applying some of the defense improvements to GPT-6 Astra as well.
The caveats: this is the company's own internal evaluation, and it measures resistance to known classes of attacks, not future ones.
OpenAI also frames distillation as a national security issue: a foreign actor that reproduces the capabilities of an American model could deploy them without safeguards. In February, the company briefed Congress on signs that DeepSeek was attempting to distill American frontier models.
The practical takeaway for anyone serving models through an API: check for reasoning leaks in messages and in tool arguments, do not allow encrypted reasoning to be reused across conversations and models, keep discovered attacks as regression tests, and supplement request-level defenses with account-level campaign detection. A model's raw reasoning is worth more than almost anything else your API returns, so treat it as sensitive data.
- Nov 20, 2026METR Updates Its Famous AI Chart: Task Length Doubles Every 4 Months, but Judgment Lags
- Nov 20, 2026AI assembles molecules with 95.4% of points, but scores only about 50% on choosing drug candidates
- Nov 20, 2026Nvidia Is Close to Buying Inception for $500 Million to Get Diffusion LLMs
- Nov 20, 2026In China, Job Postings for AI-Exposed Professions Fell 10 Percentage Points Faster Than the Rest
Key Questions Answered
Why steal "thoughts" instead of answers?
Distillation is an ordinary technique in itself: one model (the "student") learns from the data of another (the "teacher"), for example from its answers, its reasoning, or its ratings of which answers are better. Problems begin when this is done without permission. Beyond intellectual property concerns, the student picks up the capabilities but does not inherit the safeguards.
How the defense works?
The defense works on two levels. The first is the request level: the model is fine-tuned on real successful attacks from internal automated red-teaming, investigations and reports from external researchers: it refuses to reveal protected reasoning but still explains its solutions normally when the request is legitimate; input classifiers catch attempts to extract reasoning, while output classifiers look for leaks in assistant messages and in tool-call arguments, in both streaming and non-streaming responses.
What the measurements showed?
The check ran as an adaptive black-box test: attackers see the system the way an outside user does and adjust their techniques to its responses. This finds weaknesses that static tests, with a fixed set of old attacks, miss. The result is averaged across six attacker models. For extracting reasoning from a previous request, success fell from 54.0% for GPT-6 Astra to 0.2% for GPT-6.2 Astra. For the current request, it fell from 22.9% to 0.9%.
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.