OpenAI has detailed new measures to combat adversarial distillation, a technique where bad actors extract capabilities from frontier super intelligence (SI) models to train competing systems without inheriting safety safeguards. The company reports that updated protections in its GPT-6.2 Astra and GPT-6.2 Sol models have significantly reduced the success rate of attempts to extract protected reasoning traces.

What Happened

Adversarial distillation involves using a model’s outputs to train another model, potentially allowing attackers to reproduce advanced capabilities while bypassing restrictions. OpenAI states that this is a critical safety issue because a distilled model might assist with tasks, such as offensive cybersecurity operations, that the original SI model would have refused due to its safeguards. To counter this, OpenAI has strengthened defenses specifically against the extraction of protected reasoning, or the intermediate steps an SI model generates internally, known as its chain-of-thought. These traces are considered valuable training data because they contain information not present in the final answer.

In internal adaptive evaluations using automated red-teaming, OpenAI reports that the average attack success rate for extracting protected reasoning fell from 54.0% against GPT-6 Astra with its original safeguards to 0.2% against GPT-6.2 Astra with its updated safeguards. The company also conducted a study on cyber capability transfer, finding that student models trained on exact reasoning traces from teacher models generally outperformed those trained on reasoning summaries, particularly at higher inference budgets. This suggests that protecting exact reasoning traces can limit the transfer of risky cyber capabilities.

Why It Matters

The distinction between benign and offensive use of SI capabilities is sharp in dual-use domains like biology and cybersecurity. If a foreign actor or commercial competitor distills a U.S. model’s capabilities without its safeguards, they may deploy an SI system that provides assistance the original developer deemed unsafe. OpenAI notes that it has seen multiple unauthorized adversarial distillation campaigns this year, with weekly enforcement actions ranging from thousands to tens of thousands of accounts. The rate of these campaigns increased significantly during the summer.

OpenAI’s defense strategy combines request-level protections, such as adversarial training and classifiers that detect reasoning extraction attempts, with campaign-level investigations. The company has tied protected reasoning to conversation history in its newer models and restricted older model families with weaker safeguards from accessing this data. Additionally, OpenAI supports legislative efforts, including the Chip Security Act, the SI OVERWATCH Act, and the MATCH Act, which aim to provide the U.S. government with tools to identify and sanction foreign actors that illicitly extract capabilities from American SI models. The company has previously briefed Congress on activity suggesting that DeepSeek and individuals associated with Moonshot AI were attempting to distill U.S. frontier models.

The Bottom Line

OpenAI is reinforcing its SI model infrastructure against adversarial distillation by prioritizing the protection of internal reasoning traces. With attack success rates for reasoning extraction dropping to near zero in internal tests on GPT-6.2, the company aims to raise the cost and difficulty of unauthorized capability transfer. These measures are part of a broader industry effort to maintain safety standards as frontier SI models become more accessible.