# OpenAI Combatting Adversarial Distillation — new API safeguards for GPT-6.2 reasoning

> OpenAI says its GPT-6.2 models now hide their reasoning far better from copycat training.

- Org: OpenAI
- Date: 2026-11-20
- Category: security, ecosystem
- Source: https://openai.com/index/combatting-adversarial-distillation
- Canonical page: https://ai-tldr.dev/releases/openai-combatting-adversarial-distillation/

## In short

OpenAI describes new safeguards that stop other labs from copying GPT-6.2 Astra and Sol reasoning through the API. It reports attack success falling from 54.0% to 0.2% and says older model families can no longer read the reasoning.

## What it is

OpenAI's adversarial distillation post lists the safeguards it added so that GPT-6.2 Astra and Sol reasoning cannot be harvested to train rival models. They include input and output classifiers, adversarial training, and changes to how the API handles protected reasoning. OpenAI also names a Moonshot AI campaign it disrupted and shares what it learned with other providers.

## How it works

Protected reasoning is tied to the conversation that produced it: if an earlier message is edited, the reasoning after it is dropped. Older model families are blocked from reading GPT-6.2 reasoning, and a path that could expose another user's encrypted reasoning was closed. Output classifiers check assistant messages and tool-call arguments for leaks, in both streaming and non-streaming replies. OpenAI measures the result with automated red-teaming.

## Why it matters

Developers who build on the API should expect that changing earlier messages in a long agent run removes the reasoning that followed. OpenAI reports adaptive attacks that extract reasoning from an earlier request fell from 54.0% on GPT-6 Astra to 0.2% on GPT-6.2 Astra. The post also backs a consistent-standards policy for providers that offer OpenAI models.

## Who it's for

API developers and AI security teams

## Try it

Read the API guide on keeping the message prefix unchanged: https://developers.openai.com/api/docs/guides/keep-prefix-unchanged

## Quick facts

- Maker: OpenAI
- Protects: GPT-6.2 Astra and Sol reasoning
- Extraction success (earlier request): 54.0% to 0.2%
- Extraction success (current request): 22.9% to 0.9%
- API change: Editing a message drops the reasoning after it
- Enforcement: Thousands to tens of thousands of accounts weekly

## FAQ

### Will editing earlier messages break my GPT-6.2 agent runs?

OpenAI says that when an earlier message in a conversation is edited, the protected reasoning that came after it is dropped. Developers who rewrite history in long GPT-6.2 agent sessions should keep the message prefix unchanged, as OpenAI's API guide advises, or the model will lose its earlier reasoning.

### Which labs does OpenAI name in the distillation post?

OpenAI names Moonshot AI as the source of a core part of a campaign it disrupted, and says it briefed Congress about DeepSeek distillation activity. It also cites Anthropic's own distillation-defense work and shares indicators of compromise through the Frontier Model Forum. OpenAI does not publish the full evidence in the post.

### How well do the new safeguards work according to OpenAI?

OpenAI reports adaptive attacks that extract reasoning from an earlier request dropped from 54.0% on GPT-6 Astra to 0.2% on GPT-6.2 Astra. For the current request the rate fell from 22.9% to 0.9%. These are OpenAI's own red-team numbers and have not been independently reproduced.

## Links

- [OpenAI Announcement](https://openai.com/index/combatting-adversarial-distillation)
- [API Prefix Guide](https://developers.openai.com/api/docs/guides/keep-prefix-unchanged)
- [Anthropic Defenses](https://www.anthropic.com/news/distillation-defenses)
- [GPT-6.2 Launch](https://openai.com/index/introducing-gpt-6-2-astra-sol)
