---
title: OpenAI says new safeguards cut protected-reasoning extraction from 54% to 0.2% in GPT-6.2 Astra as it details its defenses against adversarial distillation | ctxsec
url: https://ctxsec.com/items/openai-combatting-adversarial-distillation-protected-reasoning-extraction-54-to-0-2-gpt-6-2-astra-classifiers-api
description: "OpenAI describes adversarial distillation as using a frontier model's outputs without permission to train a copy that lacks its safeguards, for example to rebuild cyber capabilities the original would refuse. Its defenses combine adversarial training so the model refuses to reveal protected reasoning, input classifiers for extraction attempts, output classifiers that also scan tool-call arguments, API rules that drop reasoning items when the user, organisation, safety identifier, model family or conversation history changes, and campaign-level investigations. In OpenAI's own adaptive red-team test, extracting reasoning from an earlier turn fell from 54.0% against GPT-6 Astra to 0.2% against GPT-6.2 Astra."
status: "official (OpenAI post, Nov 20, 2026; success rates are OpenAI's internal evaluation)"
published: 2026-11-20
---

> OpenAI describes adversarial distillation as using a frontier model's outputs without permission to train a copy that lacks its safeguards, for example to rebuild cyber capabilities the original would refuse. Its defenses combine adversarial training so the model refuses to reveal protected reasoning, input classifiers for extraction attempts, output classifiers that also scan tool-call arguments, API rules that drop reasoning items when the user, organisation, safety identifier, model family or conversation history changes, and campaign-level investigations. In OpenAI's own adaptive red-team test, extracting reasoning from an earlier turn fell from 54.0% against GPT-6 Astra to 0.2% against GPT-6.2 Astra.

[← AI](/ai) · [markdown](/items/openai-combatting-adversarial-distillation-protected-reasoning-extraction-54-to-0-2-gpt-6-2-astra-classifiers-api.md)

official (OpenAI post, Nov 20, 2026; success rates are OpenAI's internal evaluation)

2026-11-20 3:49 PM ET

### OpenAI says new safeguards cut protected-reasoning extraction from 54% to 0.2% in GPT-6.2 Astra as it details its defenses against adversarial distillation

- **Why defenders care:** a distilled model without safeguards can give attackers capabilities the original refuses to provide.
- **Tool-call leaks:** OpenAI notes a model can leak hidden reasoning through tool arguments, not just its answer. If you log or forward tool calls, treat their contents as sensitive.
- **Developer impact:** starting with 6.2, reasoning items are removed if the conversation prefix or `safety_identifier` changes. Apps that rewrite history may see reasoning dropped.
- The figures come from OpenAI's internal tests and have not been independently verified.

Sources: [OpenAI: Combatting adversarial distillation (Nov 20, 2026)](https://openai.com/index/combatting-adversarial-distillation/) · [OpenAI API docs: Reasoning models (reasoning item removal safeguards)](https://developers.openai.com/api/docs/guides/reasoning)
