OpenAI has published details of how it detects and disrupts adversarial distillation, in which an actor uses a frontier model's outputs without authorization to train a second model and then deploys it without the original's protections. The company says the risk is most acute in dual-use fields such as cybersecurity and biology, where a model that helps defenders could also help attackers. OpenAI has focused on protecting its models' chain-of-thought, the intermediate reasoning a model generates while working toward an answer, because those traces can include information the final answer omits. In an internal adaptive evaluation using strong automated red-teaming, 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. The company says GPT-6.2 Sol has the same strengthened defenses. In a separate internal study, OpenAI tested how much additional cyber capability transfers when a student model learns from a teacher's exact reasoning traces, compared with a baseline of content normally visible to users, including final responses, high-level reasoning outputs and tool interactions. Each student started from the same pretrained model and was trained mainly on trajectories from Astra-level teachers working on coding and defensive cybersecurity tasks. After supervised fine-tuning, the student trained on exact traces generally outperformed the student trained on the baseline content. After a lightweight round of reinforcement learning, the exact-trace student still scored higher at the largest inference budget tested on both benchmarks, while the baseline-trained student scored higher at lower budgets. OpenAI concludes that protecting reasoning traces can limit cyber capability transfer, even when the underlying task is permitted under its usage policies, and says it will keep studying how much capability transfers across student models and training methods. OpenAI describes its defenses as operating at two levels. At the request level, model behavior, classifiers and API controls limit reasoning trace extraction. At the campaign level, the company looks for patterns in request and account signals that suggest coordinated extraction, and attacks uncovered in investigations are fed back as examples for training and testing its safeguards. The company says these updated safeguards now align more closely with those it uses for biology and cybersecurity. For Australian businesses and professionals, the change is unlikely to affect everyday use of OpenAI products under standard terms, though heavier API verification and restricted access to reasoning content could affect some workflows. Anyone using model outputs to train or fine-tune their own models should check the provider's terms first, as unauthorized distillation is now being actively policed. The announcement does not alter obligations under the Privacy Act 1988, which matter only where outputs involve personal information you handle, and it does not announce any new costs.