OpenAI says distillation attacks on protected reasoning fell to 0.2%
OpenAI says extraction of protected reasoning succeeded 0.2% of the time against GPT-6.2 Astra, down from 54%.

· Last verified: 22 Nov 2026 (OpenAI)

Summary
- OpenAI says extraction of protected reasoning succeeded 0.2% of the time against GPT-6.2 Astra, down from 54%.
- Updated safeguards add adversarial training, input and output classifiers, and account-level API restrictions.
- Rival labs building models on OpenAI outputs now face refusals, detection and network-wide enforcement.
The latest
Attempts to pull protected chain-of-thought reasoning out of OpenAI's newest model succeeded 0.2% of the time in internal adaptive red-teaming, down from 54.0% against GPT-6 Astra with its original safeguards, the company said in a November 20, 2026 post. For extraction of the current request, the rate fell from 22.9% to 0.9%.
Details
- The numbers: Per-attack-type success against GPT-6.2 Astra stayed below 1%, aggregated across six attacker models, according to OpenAI. The 54.0%-to-0.2% figure covers extraction from a prior request; the current-request measure moved from 22.9% to 0.9%. The evaluation was run internally with automated adversarial red teams.
- The safeguards: OpenAI said the protections combine adversarial training, so models recognize extraction attempts and refuse to disclose protected reasoning, with classifiers on inputs that flag such attempts and classifiers on outputs that detect leakage in assistant messages and tool-call arguments, across streaming and non-streaming responses.
- API-level changes: Protections were tightened across user, workspace, organization and model-family boundaries, the company said, including closing a path that could have exposed encrypted reasoning to another user. OpenAI did not say whether that path had been exploited before it was shut.
- Conversation binding: In GPT-6.2 Astra and Sol, protected reasoning is bound to the conversation history that produced it: if an earlier message is rewritten, the API drops the reasoning that followed. Older model families with weaker safeguards are blocked from reading protected reasoning from those two models.
- Backport: OpenAI said it is applying selected improvements to GPT-6 Astra as well, extending protections beyond the newest releases. It did not set a timeline for completing that work or name which improvements are being carried over.
- Cyber study: An internal study compared a student model trained on detailed reasoning traces against one trained on reasoning summaries, using Astra-level teacher models and supervised fine-tuning data drawn mostly from coding and defensive cyber trajectories, followed by a light reinforcement-learning round.
- The findings: After supervised fine-tuning the detailed-trace student was generally stronger, OpenAI said. After the RL round it stayed ahead at the largest inference budget tested on ExploitBench and SEC-bench Pro, while the summary-trained student did better at lower budgets. OpenAI published charts rather than numerical results.
- Enforcement: OpenAI said it investigates accounts behind unauthorized extraction campaigns and acts against the surrounding account network, with weekly actions ranging "from thousands to tens of thousands of accounts," and a sharp increase last summer. It shares indicators of compromise with other model providers.
- Prior cases: The company pointed back to a September disclosure of a disrupted coordinated distillation campaign attributed largely to individuals linked to Moonshot AI, developer of Kimi, and to February congressional testimony concerning DeepSeek.
- Policy hooks: OpenAI named three pieces of legislation by title — the Chip Security Act, the AI OVERWATCH Act and the MATCH Act — and said it shares findings with industry and government. It said it plans to share more soon about recent campaigns.
Background
Distillation trains a cheaper model on a stronger model's outputs. Reasoning traces are the high-value target because they expose how a model solves a problem, not just the answer, which is why OpenAI treats chain-of-thought as protected rather than user-visible output.
Between the lines
The measured gap between a full-trace student and a summary-trained one is the commercial argument here: if detailed reasoning only wins at the largest inference budget, OpenAI's own study suggests the margin extraction buys is narrower than assumed. The enforcement posture, from account-network takedowns to shared indicators of compromise and named legislation, reads as a push to make distillation a policy problem rather than purely a technical one.
What's next
OpenAI said it plans to disclose more about recent extraction campaigns. Watch for named attribution in that disclosure, movement on the Chip Security Act, AI OVERWATCH Act and MATCH Act, and whether the backport to GPT-6 Astra is completed.
Source: OpenAI