OpenAI Says New Safeguards Cut Reasoning Extraction From 54.0% to 0.2%
A.I. / news
OpenAI Says New Safeguards Cut Reasoning Extraction From 54.0% to 0.2%
The figures are from OpenAI's own red-team evaluation, and the post names DeepSeek and Moonshot AI as distillers of U.S. models.
OpenAI said Friday that an update to its safeguards cut the rate at which automated attackers can recover its models' hidden reasoning from 54.0 percent to 0.2 percent, and that its weekly enforcement actions against distillation campaigns have ranged from thousands to tens of thousands of accounts. The figures come from OpenAI's own post, "Combatting adversarial distillation", dated Nov. 20.
Adversarial distillation means training a second model on a frontier model's outputs without permission. OpenAI's worry is the copy keeps the capability and drops the safeguards, with cyber and biology as the examples it gives.
The numbers are OpenAI's own red-team results
The test was an internal adaptive evaluation using automated red-teaming. For reasoning from an earlier prompt, the average attack success rate fell from 54.0 percent on GPT-6 Astra with the original safeguards to 0.2 percent on GPT-6.2 Astra with updated ones. For reasoning from the current request, it fell from 22.9 percent to 0.9 percent.
| Attack target | GPT-6 Astra | GPT-6.2 Astra |
|---|---|---|
| Reasoning from an earlier prompt | 54.0% | 0.2% |
| Reasoning from the current request | 22.9% | 0.9% |
OpenAI said it is applying selected improvements to GPT-6 Astra too. No outside party is named as having run the test, so both columns are vendor-supplied.
What the defences do
At the request level, OpenAI lists adversarial training, classifiers on inputs, outputs and tool-call arguments, and API restrictions. Protected reasoning is tied to the conversation that produced it and is dropped if an earlier message is rewritten. Older model families are blocked from reading GPT-6.2 Astra's reasoning.
That last rule matches the method OpenAI described on Sept. 30 in its account of a disrupted campaign. Operators there copied encrypted reasoning out of one conversation and asked a model in another conversation to decrypt and transcribe it. OpenAI said they did not break its encryption or reach stored conversations.
At the campaign level, the company says it investigates coordinated accounts, enforces against account networks and shares indicators with other providers.
Why OpenAI says hidden reasoning is worth protecting
OpenAI compared "student" models trained on exact reasoning traces with students trained on reasoning summaries. After supervised fine-tuning, the exact-trace student generally did better. After reinforcement learning, it scored higher at the largest inference budget on ExploitBench and SEC-bench Pro, while the summary-trained student scored higher at lower budgets. OpenAI concludes that protecting traces can limit how much cyber capability transfers. It cites an Anthropic result on SWE-bench Verified that showed a similar advantage.
Who is named
OpenAI said it briefed Congress in February on activity indicating that DeepSeek attempted to distill U.S. frontier models. It repeated its September attribution of "a core cluster of the activity" in the reasoning-extraction campaign to individuals associated with Moonshot AI, the developer of Kimi. The September post said it was unclear whether all operators were one actor, and the figures there counted attempts. Those were 16,000 extraction attempts from more than 4,000 users on July 24 and 25, and the campaign was fully disrupted by July 28.
The Terminal found no response from DeepSeek or Moonshot in the sources it read. The post gives no count of accounts banned over the year.
The policy section is explicit. OpenAI says it supports legislation giving the U.S. government more tools to sanction foreign actors who extract American AI capabilities, along with the Chip Security Act, the AI OVERWATCH Act and the MATCH Act.
Anthropic, whose SWE-bench Verified result OpenAI cites and which priced its IPO at $83 a share, is the only other lab the post names on the results side. For another Chinese developer's weights release, see the Terminal's report on Tencent's Hy4-flash. OpenAI says it withheld exploitable details and shared findings through the Frontier Model Forum and government channels. It does not say when the updated safeguards reach GPT-6 Astra.
Sources
More in A.I.
- 01Frostbite, an AI-Designed CPU, Claims 4.59x Efficiency Over an Unnamed BaselineArchitect Labs reports pre-silicon results from a 7 nm layout and has not said the chip has been fabricated.
- 02European Court of Human Rights Bars Romanian Lawyer for Two Years Over AI CitationsThe Court's own database could have caught the errors, according to two Romanian outlets, and sources differ on the lawyer's pronoun.
- 03NVIDIA Nemotron 3.5 Bolt Is Called 4GB; Its 4-Bit File Is 6.18 GBThe Gemma 4 E4B derivative wins seven of eight benchmark rows against its base on NVIDIA's own harness, but loses five of eight to Qwen3.5 4B.
- 04Chrome Built a Fake Web to Test Auto Browse Against Prompt InjectionTwo Chrome AI Security engineers describe a gVisor sandbox that serves generated sites at real URLs, with injections chosen by a request header.