OpenAI described on November 20, 2026 how it detects and disrupts adversarial distillation—the process by which one model’s outputs are used to train another model that may lack the original safeguards. The company emphasized protecting “protected reasoning,” the internal chain-of-thought traces that can contain details omitted from final answers.
How distillation can transfer risky capabilities
OpenAI explained that distillation allows a student model to learn from a teacher model’s answers, reasoning traces or judgments. In dual-use areas such as cybersecurity and biology, an unauthorized student could inherit capabilities while escaping safety controls that block harmful outputs.
The company reported an internal study of cyber distillation that compared students trained on exact reasoning traces to students trained on reasoning summaries. Both groups started from the same pretrained student model and were evaluated after supervised fine-tuning and a lightweight round of reinforcement learning. OpenAI said the student trained on exact traces generally outperformed the summary-trained student, and that the exact-trace student retained an advantage at the largest inference budget tested while the summary-trained student performed better at lower budgets.
Defenses and measurements
OpenAI said its approach combines adversarial testing, model training, input and output classifiers, API-level controls, and campaign-level investigations. Defensive measures include adversarial training to teach models to refuse to disclose protected reasoning, classifiers to detect extraction attempts and leaks in assistant messages and tool-call arguments, and API protections that tie protected reasoning to the conversation that produced it.
In an internal adaptive evaluation using automated red-teaming, OpenAI reported that the average attack success rate (ASR) for extracting protected reasoning fell from 54.0% on GPT‑6 Astra with original safeguards to 0.2% on GPT‑6.2 Astra with updated safeguards. For extracting reasoning from the current request, ASR fell from 22.9% to 0.9%.
OpenAI also described campaign detection and enforcement, saying weekly enforcement actions have ranged from thousands to tens of thousands of accounts. The company reported briefing Congress in February on activity linked to DeepSeek and said it attributed part of a September campaign to individuals associated with Moonshot AI, developer of Kimi.
OpenAI said it supports legislation including the Chip Security Act, the AI OVERWATCH Act, and the MATCH Act, and that selected safeguard improvements are being applied to GPT‑6 Astra while further studies of capability transfer continue.
Original source: OpenAI News