
News
OpenAI says GPT-6.2 safeguards cut reasoning extraction
OpenAI says updated GPT-6.2 Astra safeguards cut protected-reasoning extraction to 0.2% in an internal test, down from 54% on GPT-6 Astra.
OpenAI says updated safeguards cut protected-reasoning extraction on GPT-6.2 Astra to 0.2% in an internal test, down from 54.0% on GPT-6 Astra, and that on GPT-6.2 Astra and Sol the API now binds that reasoning to the conversation that produced it.
In an adaptive black-box evaluation aggregated across six attacker models, the average success rate for extracting 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 with the updated safeguards. Extraction from the current request fell from 22.9% to 0.9%. OpenAI published the comparison on November 20, 2026, in “Combatting adversarial distillation.” Black-box means the testers had the same view an external user would have, and the automated red-teamers could change tactics as the system responded. OpenAI says the setup closely mirrors production deployments. The figures are the company’s own measurements, not an independent audit.
The updated defenses cover GPT-6.2 Astra and GPT-6.2 Sol. On GPT-6.2 Astra, both attack types in that adaptive test were below 1%. OpenAI says it is also applying selected safeguard improvements to GPT-6 Astra. The published table compares the two Astra defenders only, so the 0.2% and 0.9% rates are not a separate Sol result.
With GPT-6.2 Astra and Sol, the API ties a model’s protected reasoning to the conversation history that produced it. As new messages are added, the model 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 are prevented from reading GPT-6.2 Astra and Sol protected reasoning. OpenAI’s September disclosure described attacks that moved encrypted reasoning from one conversation into another and asked the model to reveal the reasoning in user-readable form. The company says it strengthened protections across user, workspace, organization, and model-family boundaries, including closing a pathway that could expose another user’s encrypted reasoning.
Protected reasoning is the chain-of-thought generated internally while the model works toward an answer. OpenAI says those traces can contain information the final answer leaves out, which makes them valuable training data for a second model. Adversarial distillation, in the company’s framing, is unauthorized use of a frontier model’s outputs to train another model with similar capabilities and then deploy it without comparable protections. One example it gives is distilling cyber capabilities and using the resulting model for attacks the original model would have refused, or that its safeguards would have blocked. OpenAI says the updated safeguards align more closely with those it already uses for biology and cybersecurity.
Why exact traces show up in coding and cyber work
OpenAI’s models help with coding and defensive cybersecurity. The post’s concern is that a student trained on many of those allowed outputs can still pick up skills that apply to offensive tasks. In an internal distillation study, the company started each run from the same pretrained student. The supervised fine-tuning set consisted primarily of trajectories from Astra-level teachers working on coding and defensive cybersecurity tasks. Each trajectory included the steps the teacher took, any tool interactions, and the final response.
One student learned from content normally visible to users: final responses, reasoning summaries, and tool interactions. The other had access to the teacher’s exact reasoning traces. After supervised fine-tuning, the exact-trace student generally scored higher. After a lightweight round of reinforcement learning in cybersecurity and coding environments, the exact-trace student still scored higher at the largest inference budget tested on both benchmarks shown, while the summary-trained student scored higher at lower budgets. One benchmark in the post is labeled ExploitBench. The chart compares capability coverage as the inference budget, measured in weighted output tokens, increases. Dashed lines are after fine-tuning; solid lines are after the further reinforcement-learning round.
OpenAI reads the result as a reason to protect exact traces even when the teacher’s task is allowed under its usage policies. Across many interactions, it says, those traces can teach skills that generalize beyond the original tasks. The company says it will keep studying how much additional capability transfers across student models and distillation recipes. A footnote points to related Anthropic work that reported a similar advantage for full reasoning transcripts over summaries on SWE-bench Verified, and says OpenAI’s result shows the advantage extends to frontier cybersecurity. Like the attack-success rates, this is an internal study, not an outside reproduction.
How the request and campaign layers fit together
Request-level controls are meant to limit extraction in each interaction. Adversarial training teaches models to recognize reasoning-extraction attempts and refuse to disclose protected reasoning, while still answering legitimate requests for explanations. Examples come from internal automated red-teaming, investigations, and responsible disclosures. Input classifiers look for extraction attempts. Output classifiers look for leaks in assistant messages and tool-call arguments, on streaming and non-streaming responses. OpenAI says tool-call coverage matters because a model can expose protected reasoning in a tool’s arguments without including it in the final answer, and that output classifiers can block a suspected leak before it reaches the caller.
Campaign detection looks across requests and accounts for coordinated extraction. When OpenAI says it identifies violations, it enforces against the network of accounts. Findings from those investigations, it says, become training and test examples for the safeguards. Weekly enforcement actions against unauthorized adversarial distillation campaigns have ranged from thousands to tens of thousands of accounts, with the rate increasing significantly “this past summer.” The post does not connect that range to the GPT-6.2 extraction test. OpenAI says it has seen multiple such campaigns on its platform this year, shared initial findings in a September report, and shares indicators of compromise with other model providers.
The same post treats distillation as a national-security issue. OpenAI says it briefed Congress in February on activity indicating that DeepSeek was attempting to distill U.S. frontier models, and that its September report attributed a core part of a protected-reasoning extraction campaign to individuals associated with Moonshot AI, the developer of Kimi. It says it plans to share more soon about recent campaigns it has disrupted. It also says it has told Congress it supports the Chip Security Act, the AI OVERWATCH Act, and the MATCH Act. Its argument is that export controls and distillation defenses are complementary: one limits access to the hardware needed to train frontier models, and the other makes it harder to acquire some of the same capabilities from models already trained.
If you call GPT-6.2 Astra or Sol, check what your client does when it rewrites an earlier message: the API drops the reasoning that followed that turn, and older model families cannot read those traces. Read OpenAI’s November 20 post for the attack-success table and the coding and cyber comparison, and treat the 0.2% and 0.9% figures as the company’s internal adaptive results until they are reproduced outside OpenAI.
Sources
- Combatting adversarial distillation openai.com