OpenAI Locks Up Its Reasoning Traces: Extraction Success Falls From 54% to 0.2%
Chain of thought is now treated as an asset to guard, not a byproduct. On Friday OpenAI published how it fights adversarial distillation: someone uses a frontier model's outputs without authorization to train another model, then ships it without comparable safeguards. OpenAI frames this as a safety problem as well as an IP one. A model distilled from a defensive cyber assistant could help attackers in ways the original would have refused. Post: https://openai.com/index/combatting-adversarial-distillation/
The headline number comes from an internal adaptive black-box evaluation, with automated red teamers adjusting their attacks across six attacker models. Success at recovering protected reasoning from an earlier request fell from 54.0% against GPT-6 Astra with its original safeguards to 0.2% against GPT-6.2 Astra. For reasoning from the current request it fell from 22.9% to 0.9%. The defenses are layered. Adversarial training teaches the model to refuse extraction attempts. Input and output classifiers also watch tool-call arguments, because reasoning can leak there without appearing in the answer. At the API level, reasoning is now tied to the conversation history that produced it: rewrite an earlier message and the reasoning after it is dropped, and older model families cannot read GPT-6.2's protected reasoning. Some of these fixes are being backported to GPT-6 Astra.
Why the exact traces matter is the most interesting part, and it is more nuanced than the framing. OpenAI started from one pretrained student and fine-tuned it on Astra-level teacher trajectories from coding and defensive cyber tasks. After supervised fine-tuning, the student trained on exact reasoning traces generally beat the one trained on the reasoning summaries users see. After a light round of RL, the exact-trace student led only at the largest inference budget, and the summary-trained student scored higher at lower budgets. The edge is real, but narrower than "the traces are the crown jewels."
The scale is not small. Weekly enforcement actions range from thousands to tens of thousands of accounts, and the rate rose sharply this summer. OpenAI recalls briefing Congress in February on DeepSeek distillation attempts, and in September attributing part of a campaign to individuals associated with Moonshot AI. It says it supports the Chip Security Act, the AI OVERWATCH Act and the MATCH Act.
My read: this creates a real tension with transparency. The traces that safety researchers, auditors and agent builders most want to read are exactly the ones being locked away. On the same day, Thinking Machines moved part of its model's reasoning into silent loops. One lab hides reasoning in architecture, another in a vault, and the result for users is the same: less of the thinking is visible. If you build on these APIs, expect logs, replays and debugging to rely more and more on summaries the vendor chooses to show you.
← Back to all articles
The headline number comes from an internal adaptive black-box evaluation, with automated red teamers adjusting their attacks across six attacker models. Success at recovering protected reasoning from an earlier request fell from 54.0% against GPT-6 Astra with its original safeguards to 0.2% against GPT-6.2 Astra. For reasoning from the current request it fell from 22.9% to 0.9%. The defenses are layered. Adversarial training teaches the model to refuse extraction attempts. Input and output classifiers also watch tool-call arguments, because reasoning can leak there without appearing in the answer. At the API level, reasoning is now tied to the conversation history that produced it: rewrite an earlier message and the reasoning after it is dropped, and older model families cannot read GPT-6.2's protected reasoning. Some of these fixes are being backported to GPT-6 Astra.
Why the exact traces matter is the most interesting part, and it is more nuanced than the framing. OpenAI started from one pretrained student and fine-tuned it on Astra-level teacher trajectories from coding and defensive cyber tasks. After supervised fine-tuning, the student trained on exact reasoning traces generally beat the one trained on the reasoning summaries users see. After a light round of RL, the exact-trace student led only at the largest inference budget, and the summary-trained student scored higher at lower budgets. The edge is real, but narrower than "the traces are the crown jewels."
The scale is not small. Weekly enforcement actions range from thousands to tens of thousands of accounts, and the rate rose sharply this summer. OpenAI recalls briefing Congress in February on DeepSeek distillation attempts, and in September attributing part of a campaign to individuals associated with Moonshot AI. It says it supports the Chip Security Act, the AI OVERWATCH Act and the MATCH Act.
My read: this creates a real tension with transparency. The traces that safety researchers, auditors and agent builders most want to read are exactly the ones being locked away. On the same day, Thinking Machines moved part of its model's reasoning into silent loops. One lab hides reasoning in architecture, another in a vault, and the result for users is the same: less of the thinking is visible. If you build on these APIs, expect logs, replays and debugging to rely more and more on summaries the vendor chooses to show you.
Comments