OpenAI Says an Internal Test Cut Earlier Reasoning Extraction From 54.0% to 0.2%
OpenAI said on November 20, 2026 that an internal adaptive test of reasoning extraction fell from 54.0 percent against GPT-6 Astra, with its original safeguards, to 0.2 percent against GPT-6.2 Astra, with updated safeguards. The test asked whether an attacker could recover protected reasoning the model had generated for an earlier prompt. A second row, extracting reasoning from the current request, fell from 22.9 percent to 0.9 percent. Both figures are OpenAI's, and they compare two model versions as well as two safeguard stacks.
The numbers
- 54.0 percent to 0.2 percent. Adaptive black-box attack success rate for extracting protected reasoning from an earlier request. GPT-6 Astra with original safeguards versus GPT-6.2 Astra with updated safeguards. Aggregated across six attacker models. OpenAI, November 20.
- 22.9 percent to 0.9 percent. Same setup, extracting protected reasoning from the current request. OpenAI says both rates on GPT-6.2 Astra were below 1 percent.
- Weekly enforcement against unauthorized distillation campaigns has ranged from thousands to tens of thousands of accounts, and OpenAI says the rate increased significantly last summer. The post does not give a week-by-week table.
What the post means by adversarial distillation
OpenAI defines adversarial distillation as using a frontier model's outputs without authorization to train another model with similar capabilities, then deploying that model without comparable safeguards. The example it gives is distilling cyber capability and using the student model for attacks the teacher would have refused. The post says the lab is most concerned with protected reasoning, the chain of thought a model generates internally, because those traces can hold information the final answer leaves out.
GPT-6.2 Astra and GPT-6.2 Sol shipped earlier this month. This November 20 post is the safeguard note, not the model release. OpenAI says it is applying selected safeguard improvements to GPT-6 Astra as well. It does not say those improvements are finished, and it does not reprint the 0.2 percent figure for GPT-6 Astra.
What changed in the API
The post says GPT-6.2 Astra and GPT-6.2 Sol tie protected reasoning to the conversation history that produced it. New messages can build on earlier reasoning. If an earlier message is rewritten, the API drops the reasoning that followed it. Older model families with weaker safeguards cannot read GPT-6.2 Astra and Sol protected reasoning. OpenAI says a September disclosure described attacks that moved encrypted reasoning from one conversation into another. Input classifiers look for extraction attempts. Output classifiers look for leaks in assistant messages and in tool-call arguments, on streaming and non-streaming responses.
The cyber study does not print a score
In an internal study, OpenAI trained student models on coding and defensive cybersecurity trajectories from Astra-level teachers. One student saw the content users normally see, including reasoning summaries. The other saw exact reasoning traces. After supervised fine-tuning, the exact-trace student generally outperformed the summary-trained student. After a lightweight round of reinforcement learning, the exact-trace student still scored higher at the largest inference budget tested on both benchmarks shown, and the summary-trained student scored higher at lower budgets. The post does not name those benchmarks or print the scores in the text. A footnote points to related Anthropic work on SWE-bench Verified, where full transcripts also beat summaries. That footnote is a citation, not an OpenAI score.
What is not disclosed?
The six attacker models are not named. The post does not say an outside lab reproduced the 0.2 percent figure. It does not give a count of reasoning tokens blocked, or a dollar figure for the program. On campaigns, it says OpenAI briefed Congress in February on activity indicating DeepSeek was attempting to distill U.S. frontier models, and that in September it attributed a core part of a reasoning-extraction campaign to individuals associated with Moonshot AI, the developer of Kimi. Those are prior disclosures this post repeats. It says more on recent campaigns is coming, and it does not name a new actor. It says OpenAI supports the Chip Security Act, the AI OVERWATCH Act, and the MATCH Act. It does not quote bill text.
Quick answers
- What happened with OpenAI?
- OpenAI said on November 20, 2026 that an internal adaptive test of reasoning extraction fell from 54.0 percent against GPT-6 Astra, with its original safeguards, to 0.2 percent against GPT-6.2 Astra, with updated safeguards. The test asked whether an attacker could recover protected reasoning the model had generated for an earlier prompt. A second row, extracting reasoning from the current request, fell from 22.9 percent to 0.9 percent.
- What are the key numbers?
- 54.0 percent to 0.2 percent. Adaptive black-box attack success rate for extracting protected reasoning from an earlier request. GPT-6 Astra with original safeguards versus GPT-6.2 Astra with updated safeguards. Aggregated across six attacker models. OpenAI, November 20.
- What is not disclosed?
- The six attacker models are not named. The post does not say an outside lab reproduced the 0.2 percent figure. It does not give a count of reasoning tokens blocked, or a dollar figure for the program. On campaigns, it says OpenAI briefed Congress in February on activity indicating DeepSeek was attempting to distill U.S.
This article was ultrathought.