Skip to content
AI Security Wire

Published

- 5 min read

By

OpenAI Details Anti-Distillation Defenses for Protected Reasoning

img of OpenAI Details Anti-Distillation Defenses for Protected Reasoning

OpenAI published a post on 20 November 2026 describing how it detects and disrupts adversarial distillation, and how it hardened what it calls protected reasoning, the intermediate steps a model works through before answering. The figures below are OpenAI’s own, from its internal testing; no independent replication is reported in the post; its footnote cites outside work on a related finding (SWE-bench Verified), not the 0.2% result.

What OpenAI says changed

OpenAI says it strengthened defenses against protected reasoning extraction in GPT-6.2 Astra and GPT-6.2 Sol. In an internal adaptive evaluation using automated red-teaming, it tested whether attackers could recover reasoning the model had generated while answering an earlier prompt. The average attack success rate fell from 54.0% against GPT-6 Astra with its original safeguards to 0.2% against GPT-6.2 Astra with updated safeguards. For extraction from the current request, the figures were 22.9% and 0.9%. OpenAI aggregated these across six attacker models in black-box testing, and says it is now applying selected improvements to GPT-6 Astra as well.

Three layers are described at the request level:

  • Model training: adversarial training on successful attacks found through red-teaming, investigations and outside disclosures, intended to refuse requests to reveal protected reasoning while still answering legitimate requests for explanations.
  • Classifiers: input classifiers for extraction attempts, and output classifiers for leaks in assistant messages and in tool-call arguments, applied to streaming and non-streaming responses. OpenAI notes a model can expose protected reasoning in a tool’s arguments without putting it in the final answer.
  • API protections: with GPT-6.2 Astra and Sol, protected reasoning is tied to the conversation history that produced it. If an earlier message is rewritten, the API drops the reasoning that followed. OpenAI also says older model families with weaker safeguards cannot read the new models’ protected reasoning, and that it closed a pathway that allowed someone who already possessed another user’s encrypted reasoning to replay it and recover its contents.

The cyber transfer study

OpenAI also reports an internal study on whether exact reasoning traces transfer more capability than the reasoning summaries users normally see. It fine-tuned students from the same pretrained model on trajectories from Astra-level teachers working on coding and defensive security tasks. After supervised fine-tuning, the exact-trace student generally outperformed the summary-trained student. After a lightweight reinforcement learning round, the exact-trace student still scored higher at the largest inference budget tested on both benchmarks shown, while the summary-trained student scored higher at lower budgets. OpenAI says it will keep studying how far this transfers across other student models and training recipes.

Campaigns and enforcement

OpenAI says it has seen multiple unauthorized distillation campaigns this year, with weekly enforcement actions ranging from thousands to tens of thousands of accounts and a rate that rose significantly over the summer. It says it shares indicators of compromise with other model providers.

The post revisits OpenAI’s September report, which described a campaign beginning on 1 July in which operators copied encrypted reasoning from one conversation and asked a model in another to decrypt and transcribe it. That report cited spikes on 24 and 25 July of 16,000 requests from over 4,000 users, and a related cluster of more than 15,000 users disrupted by 28 July. OpenAI attributed a core cluster to individuals associated with Moonshot AI, the developer of Kimi, while saying it was unclear whether all operators were a single actor. The new post adds that OpenAI briefed Congress in February on activity indicating DeepSeek was attempting to distill US frontier models, and that it plans to share more on recent campaigns. We found no new attributions in this post.

Why this matters for defenders

OpenAI says that beyond intellectual property concerns, unauthorized distillation poses safety risks, most clearly in dual-use fields: a student trained on benign coding and defensive security outputs may pick up skills that also apply offensively, without the teacher’s refusals. OpenAI says its results suggest that protecting reasoning traces can limit cyber capability transfer, even when the underlying task is allowed under its usage policies. That is OpenAI’s own interpretation of an internal study that used its own models and benchmarks.

Our assessment, not OpenAI’s: teams that operate model APIs should treat reasoning content and tool-call arguments as leak channels, and test them. OpenAI’s own advice to other developers matches that: check assistant messages and tool-call arguments for leaks, restrict reuse of encrypted reasoning across conversations and models, keep successful extraction attacks as regression tests, combine request-level controls with campaign-level investigation, and coordinate safeguards with deployment partners.

The post also carries a policy position. OpenAI says it supports legislation giving the US government more tools to sanction foreign actors that extract capabilities from American models, and lists three Acts it has told Congress it supports.

What is not known

The 0.2% figure comes from OpenAI testing its own defenses with its own automated attackers, and in the two posts we read we found no published attack set. In those posts we also found no naming of the campaigns behind the recent enforcement actions and no count of weekly accounts attributed to a named actor. Our reading: neither a lower success rate in a test nor a larger number of banned accounts shows how much capability has already been extracted.

References

Frequently Asked Questions

What is adversarial distillation?
OpenAI defines it as using a frontier model's outputs, reasoning or judgments without authorization to train another model with similar capabilities, which is then deployed without comparable safeguards. OpenAI frames it as a safety and national security issue, not only an intellectual property one.
How effective are the new defenses?
The figures are OpenAI's own. In its internal adaptive black-box red-teaming, aggregated across six attacker models, the success rate for recovering reasoning from an earlier request fell from 54.0% on GPT-6 Astra with original safeguards to 0.2% on GPT-6.2 Astra with updated safeguards. For the current request it fell from 22.9% to 0.9%. We found no evidence of an outside reproduction as of 21 Nov 2026 in the two OpenAI posts we read.
Does this matter if I do not run a frontier model?
Our reading: yes, for anyone exposing a model API that returns reasoning content or tool-call arguments. OpenAI's own advice to other developers is to check assistant messages and tool-call arguments for reasoning leaks, restrict reuse of encrypted reasoning across conversations and models, and keep successful extraction attacks as regression tests.