OpenAI Details How It Fights Attempts to Copy Its Models' Reasoning

OpenAI published its approach to adversarial distillation on 20 November. It says exact reasoning traces transfer more cyber skill to copycat models, and that updated safeguards held adaptive extraction attacks below 1%.

By Utopia Data & AI Team6 min read

A technician in a red plaid shirt crouches beside open server racks in a data centre, working on a laptop on the floor
Licensed photo. A technician works at a server rack in the NERSC data centre, photographed on 14 October 2011. The image is contextual and shows no OpenAI equipment. Photo: Derrick Coetzee/Wikimedia Commons, CC0.[2][1]

In brief

  • OpenAI published a description on 20 November 2026 of how it detects and disrupts adversarial distillation. In that practice, an actor uses a frontier model's outputs without permission to train another model with similar skills but weaker safeguards.[1]
  • Its study found that a student model trained on exact reasoning traces generally beat one trained on summaries after supervised training. OpenAI concludes that protecting traces can limit how much cyber skill transfers.[1]
  • OpenAI reports that adaptive extraction attacks against GPT-6.2 Astra succeeded well under 1% of the time with updated safeguards, against 22.9% to 54.0% for the original GPT-6 Astra.[1]

What OpenAI published on 20 November

OpenAI published a research and safety post on 20 November 2026 describing how it fights adversarial distillation. The company defines it as using a frontier model's outputs without authorisation to train a model with comparable capability but weaker safeguards.[1]

The post puts special weight on what OpenAI calls protected reasoning, meaning the model's chain-of-thought. It sets out a cyber-capability study, request-level and campaign-level defences, measured attack results and recommendations for other developers.[1]

This story rests on OpenAI's own post. The attack-success figures come from OpenAI's internal red-teaming, and we have not reproduced them.[1]

Rows of empty black server racks in a data-centre room with a raised white tile floor and a grey cabinet at the far wall
Licensed photo. Server racks in a data-centre room, photographed on 4 November 2015. The image is contextual and shows no OpenAI equipment. Photo: Carl Lender/Wikimedia Commons, CC BY 2.0.[3][1]

Why OpenAI protects reasoning traces

OpenAI trained student models from the same pretrained base. Most of the supervised fine-tuning data came from teacher-model trajectories on coding and defensive cybersecurity tasks, from Astra-level models.[1]

After supervised training, the student that learned from exact reasoning traces generally outperformed the student that learned from reasoning summaries. After a light round of reinforcement learning, the exact-trace student still scored higher at the largest inference budget on both ExploitBench and SEC-bench Pro. The summary-trained student scored higher at smaller budgets.[1]

OpenAI concludes that protecting traces can limit cyber capability transfer, even for tasks its usage policies allow. A footnote says Anthropic reported a similar advantage on SWE-bench Verified.[1]

Request-level and campaign-level defences

At the request level, OpenAI trains models to recognise extraction attempts and to refuse to reveal protected reasoning, while still explaining answers to legitimate users. Input classifiers flag attempts, and output classifiers catch leaks in assistant messages and tool-call arguments, in streaming and non-streaming responses.[1]

OpenAI also changed its API. After a September disclosure, it strengthened boundaries between users, workspaces, organisations and model families, and closed a path that could have exposed another user's encrypted reasoning. For GPT-6.2 Astra and GPT-6.2 Sol, reasoning is tied to the conversation history that produced it, and the API drops later reasoning if an earlier message is rewritten.[1]

At the campaign level, OpenAI investigates the accounts behind detected campaigns and acts against networks of accounts. It shares indicators of compromise with other providers and applies shared safeguard standards to providers that offer its models.[1]

Close view of the rear of a server rack with rows of small blue status screens and bundles of white Ethernet cables
Licensed photo. The rear of a server rack at the NERSC data centre, photographed on 14 October 2011. The image is contextual and shows no OpenAI equipment. Photo: Derrick Coetzee/Wikimedia Commons, CC0.[4][1]

Measured results, campaign scale and policy

In adaptive black-box red-teaming across six attacker models, OpenAI reports that attack success on the earlier request 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 says the average was below 1% for each attack type and that it is applying selected improvements to GPT-6 Astra.[1]

OpenAI says it has seen several unauthorised campaigns this year. Weekly enforcement has ranged from thousands to tens of thousands of accounts, and the rate rose significantly over the summer. It says it briefed Congress in February on activity indicating DeepSeek was trying to distill US frontier models. In September it reported disrupting a campaign it attributes in part to individuals associated with Moonshot AI. We have not verified either attribution.[1]

OpenAI backs legislation to give the US government more tools against foreign actors that extract AI capability illicitly, and names the Chip Security Act, the AI OVERWATCH Act and the MATCH Act. It urges other developers to protect private reasoning across whole interactions, keep successful attack logs as regression tests and coordinate with deployment partners.[1]

Why Malaysia should care

Our analysis: OpenAI names no Malaysian party. Malaysian companies that build products on frontier model APIs should expect tighter limits on how reasoning is returned and reused, because OpenAI says it now ties reasoning to the conversation that produced it. Teams that cache or rewrite model messages should test their integrations against these API changes.

Malaysian developers using OpenAI's API

Reasoning is now bound to its conversation history for GPT-6.2 Astra and Sol. Code that edits earlier messages may lose later reasoning.[1]

Practical move: Test any message-rewriting or caching code against GPT-6.2 models.

Malaysian AI start-ups training their own models

OpenAI treats unauthorised training on model outputs as abuse. Using outputs from any provider needs a check against that provider's terms.[1]

Practical move: Review the terms for every model output in your training data.

Security and compliance teams

OpenAI recommends logging successful extraction attacks as regression tests and coordinating with partners.[1]

Practical move: Add extraction attacks to your own red-team test suites.

What Malaysians can do now

  1. Read OpenAI's post in full before repeating the 0.2% or 0.9% figures.
  2. Test integrations that rewrite earlier messages against GPT-6.2 Astra and Sol.
  3. Check each provider's terms before using model outputs as training data.

What we still do not know

What we still do not know

  • Whether outside researchers can reproduce OpenAI's attack-success results.
  • Whether the DeepSeek and Moonshot AI attributions will be confirmed independently.
  • Whether other providers will adopt the same protections for reasoning.

Sources

  1. Combatting adversarial distillationOpenAI, 20 November 2026
  2. File:Technician with laptop working on server rack at NERSC.jpgWikimedia Commons, 14 October 2011
  3. File:Datacenter Server Racks (22370909788).jpgWikimedia Commons, 4 November 2015
  4. File:Rear of rack at NERSC data center - closeup.jpgWikimedia Commons, 14 October 2011

Facts come from OpenAI's own post dated 20 November 2026, read on 21 November. Figures and attributions are those OpenAI reported, and we did not test or audit them. The section marked as our analysis is Utopia's interpretation. The three photographs are licensed contextual images of data-centre equipment and show nothing from OpenAI. No generated image is used.

Last fact check: 21 November 2026. Factual correction? devteam.utopia@gmail.com.

More AI News

Back to AI NewsAll 215 stories