OpenAI hardens GPT-6.2 models against extraction of their hidden reasoning
OpenAI says new safeguards cut reasoning-extraction attack success on GPT-6.2 Astra to under 1% in its own tests, and its API now drops reasoning after edited messages.
OpenAI says it has tightened the safeguards that stop people from pulling hidden reasoning out of its newest models. In a post published on November 20, 2026, the company explained how it fights adversarial distillation, where someone uses a model's outputs without permission to train a rival model that lacks the original safety controls.
What's new
- Focus on chain-of-thought: OpenAI says it puts special weight on protecting a model's internal reasoning steps, because those traces can contain information missing from the final answer and make especially valuable training data.
- Measured drop in attacks: in OpenAI's internal adaptive red-team tests, the average success rate for recovering reasoning from an earlier request fell from 54.0% on GPT‑6 Astra with its original safeguards to 0.2% on GPT‑6.2 Astra. For reasoning from the current request it fell from 22.9% to 0.9%.
- Layered defences: the company describes adversarial training so models refuse extraction attempts, input and output classifiers that also watch tool-call arguments, and account-level investigations into coordinated campaigns.
- API behaviour change: with GPT‑6.2 Astra and GPT‑6.2 Sol, protected reasoning is tied to the conversation that produced it. If an earlier message is rewritten, OpenAI says its API discards the reasoning that came after it, and older model families can no longer read the new models' protected reasoning.
- Scale of enforcement: OpenAI says its weekly actions against distillation campaigns this year have ranged from thousands to tens of thousands of accounts.
- Older model: some of the improvements are now being applied to GPT‑6 Astra as well, per the post.
Why it matters
Analysis: for most users this work is invisible, but developers should note the API change. Apps that edit or replay earlier turns in a conversation may find that reasoning built up after the edited message is no longer carried forward, which can affect agent workflows that rewrite history to save tokens. More broadly, the post shows frontier labs treating reasoning traces as sensitive assets, which is one reason full chain-of-thought is rarely shown to users.
Source: OpenAI