OpenAI Details How It Detects and Disrupts Adversarial Distillation

OpenAI is describing how it detects and disrupts adversarial distillation, the practice of using a frontier model’s outputs to train another model with similar capabilities but weaker safeguards. The company focuses on protected reasoning: the internal chain-of-thought traces a model generates on the way to an answer. These traces can contain information the final answer leaves out, which makes them especially valuable training data. OpenAI says it has strengthened defenses in GPT-6.2 Astra and GPT-6.2 Sol. In an internal adaptive evaluation, the average attack success rate for recovering protected reasoning fell from 54.0% against GPT-6 Astra with its original safeguards to 0.2% against GPT-6.2 Astra with updated safeguards.

Distillation itself is a common machine-learning technique, but unauthorized distillation is a safety issue in dual-use domains such as cybersecurity and biology. A student model trained on many benign coding and defensive security outputs may learn skills that also apply to offensive tasks. To study capability transfer, OpenAI used the content normally visible to users, including final responses, reasoning summaries, and tool interactions, as a baseline, and then tested how much additional cyber capability transfers when the student has access to exact reasoning traces. Starting from the same pretrained student model, the researchers created supervised fine-tuning datasets from trajectories produced by Astra-level teachers and then applied a lightweight reinforcement learning round. After supervised fine-tuning, the student trained on exact reasoning traces generally outperformed the student trained on reasoning summaries. After reinforcement learning, the exact-trace student scored higher at the largest inference budget tested on both benchmarks shown, while the summary-trained student scored higher at lower budgets. These results suggest that protecting reasoning traces can limit cyber capability transfer, even when the underlying tasks are allowed under usage policies.

OpenAI’s defenses work at two levels: request-level safeguards and campaign-level detection. Adversarial training teaches models to recognize and refuse reasoning extraction attempts while preserving legitimate explanations. Input classifiers detect attempts to extract protected reasoning, and output classifiers detect leaks in assistant messages and tool-call arguments, including in streaming responses. API-level protections restrict how encrypted reasoning can be reused across conversations, users, workspaces, organizations, and model families. For example, if an earlier message is rewritten, the API drops the reasoning that followed it, and older model families with weaker safeguards cannot read protected reasoning produced by GPT-6.2 Astra or GPT-6.2 Sol. At the campaign level, OpenAI investigates coordinated extraction attempts, enforces against networks of accounts, and shares indicators of compromise with other model providers.

OpenAI also reports that in adaptive black-box red-teaming, attack success rates on GPT-6.2 Astra were below 1% for recovering protected reasoning from either an earlier request or the current one. Selected safeguards are now being applied to GPT-6 Astra as well. On enforcement, weekly actions against adversarial distillation campaigns this year have ranged from thousands to tens of thousands of accounts, with the rate increasing significantly over the past summer.

The company frames adversarial distillation as a national security risk because a foreign actor can reproduce U.S. frontier model capabilities and deploy them without the developer’s protections. OpenAI briefed Congress in February on activity indicating that DeepSeek was attempting to distill U.S. frontier models, and in September it reported disrupting a campaign to extract protected reasoning, attributing part of that activity to individuals associated with Moonshot AI. OpenAI says it supports legislation including the Chip Security Act, the AI OVERWATCH Act, and the MATCH Act, as well as stronger export controls.

OpenAI recommends that other model developers protect private reasoning across the full interaction, check assistant messages and tool-call arguments for leaks, keep successful attack logs as regression tests, combine request-level mitigations with campaign-level enforcement, and coordinate safeguards with deployment partners. It also says it will continue studying how much additional capability transfers across different student models and distillation training recipes.

Combatting adversarial distillation

View Original