OpenAI Details Safeguards Against Adversarial Distillation
OpenAI said updated safeguards reduced the average success rate for extracting an earlier request’s protected reasoning to 0.2% on GPT-6.2 Astra, compared with 54.0% on GPT-6 Astra under its original protections. The company published the internal evaluation results alongside analysis and practical guidance on adversarial distillation on November 20, 2026.
The figures came from an adaptive black-box red-team evaluation combining results across six attacker models. For attempts targeting reasoning from the current request, OpenAI reported success rates of 0.9% for GPT-6.2 Astra and 22.9% for GPT-6 Astra.
In OpenAI’s explanation, distillation uses one model’s output, including answers, reasoning or quality assessments, to train another model. The company said extracting a frontier model’s capabilities without authorization or comparable safeguards presents safety concerns, particularly in cybersecurity and biology, where capabilities have dual uses.
For GPT-6.2 Astra and GPT-6.2 Sol, OpenAI said its protections combine API controls, adversarial testing, training, classifiers and investigations into coordinated extraction. These measures use approaches also applied to biology and cybersecurity. Its training distinguishes requests for protected reasoning, which models are taught to refuse, from legitimate requests for explanations.
According to OpenAI, classifiers examine incoming requests for extraction attempts and check outgoing answers and tool-call arguments for leaked reasoning, whether responses are streamed or delivered without streaming. The API also ties protected reasoning to its originating conversation: editing an earlier message removes the reasoning generated after it. Older model families with weaker protections cannot access the two newer models’ protected reasoning, the company said.
OpenAI said exact internal reasoning can include intermediate information omitted from final answers. A separate internal cybersecurity-capability study compared student models sharing the same pretrained starting point. One training baseline comprised visible answers, reasoning summaries and tool interactions; the comparison data included complete task trajectories produced by Astra-level teacher models performing coding and defensive cybersecurity work.
Beyond individual requests, OpenAI said it examines account and request signals together for patterns, investigates suspected coordinated extraction and applies enforcement across account networks when it identifies violations. It also shares indicators of compromise with other model providers. During 2026, the company reported weekly enforcement against adversarial-distillation campaigns involving thousands to tens of thousands of accounts, with the rate rising during summer.
The guidance asks developers and platforms to examine tool-call arguments alongside ordinary answers for reasoning leaks and preserve successful extraction attempts as regression tests. OpenAI also recommended repeated testing against established attacks and adaptive red-teamers, along with coordination between deployment partners on safeguards and enforcement.
* This article is based on publicly available information at the time of writing.