OpenAI published on Thursday (November 20) a detailed account of how it fights adversarial distillation: the unauthorized use of a frontier model's outputs — including its internal reasoning — to train another model with similar capabilities but without the same safeguards. The company's focus is protecting so-called protected reasoning (chain-of-thought): the intermediate steps a model generates before answering, which can contain information the final answer leaves out and are therefore especially valuable training data for anyone trying to copy a model's capabilities.
From 54% to 0.2%, according to OpenAI
The headline figure comes from OpenAI's own internal evaluation, and should be read as such: in an adaptive automated red-teaming test — in which attackers adjust their attempts based on the system's responses, with the same view an external user would have — the average success rate for recovering protected reasoning from an earlier request fell from 54.0% on GPT-6 Astra, with its original safeguards, to 0.2% on GPT-6.2 Astra, with the updated safeguards. For extracting reasoning from the current request, the rate fell from 22.9% to 0.9%. The results aggregate six attacker models. These are OpenAI's measurements, from OpenAI's tests — not an independent verification.
How the defenses against adversarial distillation work
The protections operate on two levels. At the level of each request, OpenAI combines adversarial training — so the model refuses to disclose protected reasoning while still answering legitimate requests for explanations — input and output classifiers that check answers and tool-call arguments for leaks, and API controls. In GPT-6.2 Astra and GPT-6.2 Sol, protected reasoning is tied to the conversation history that produced it: if an earlier message is rewritten, the API drops the reasoning that followed, and older model families can no longer read it. At the campaign level, the company monitors request and account patterns to detect coordinated extraction, enforces bans, and shares indicators of compromise with other model providers. According to OpenAI, weekly enforcement actions against distillation campaigns range from thousands to tens of thousands of accounts, and the pace increased significantly in the summer of 2026.
The background: the Moonshot AI-linked campaign
The new publication on adversarial distillation continues an episode OpenAI had already disclosed: in July 2026, operators tried to extract protected reasoning by copying encrypted reasoning from one conversation into another and asking the model to decrypt and transcribe it. On July 24 and 25, the company observed about 16,000 requests using that pattern, from more than 4,000 users, and identified a related cluster of more than 15,000 users, disrupted by July 28. OpenAI attributes the core of the activity to individuals associated with Moonshot AI, the developer of Kimi — an attribution by the company, not a judicial finding. Geekiki covered the disruption of the campaign in October; this week's news is the defense framework and the effectiveness figures OpenAI is now disclosing. The company's conclusion is that protecting reasoning traces can limit the transfer of capabilities in sensitive areas, such as cyber operations, even when the underlying task is allowed under its usage rules.
Image: the OpenAI logo seen through a magnifying glass. Photo: Jernej Furman, CC BY 2.0, via Wikimedia Commons.
Sources
- OpenAI — Combatting adversarial distillation
- OpenAI — Disrupting a coordinated model-distillation campaign
- The Decoder — OpenAI makes it much harder for attackers to extract how its latest AI models think
- Unite.AI — OpenAI Strengthens GPT-6.2 Astra Against Reasoning Extraction
Earlier Geekiki coverage: OpenAI interrompe campanha de destilação e aponta núcleo ligado à Moonshot AI · OpenAI lança GPT-6.2 Astra e Sol com API 20% mais barata
Transparency note: this article was produced with the assistance of artificial intelligence, cross-checked against public posts on X and internet sources, and reviewed before publication. Please consult the original sources to confirm the information.
By Muse