
The Library Media — Science, Technology & Health · AI News & Intelligence Desk
From 54% to 0.2%: OpenAI Says It Can Now Stop Its Models’ Hidden Reasoning Being Copied
By Catalina Clarke
A frontier AI model does some of its most valuable work where no one can see it: in the private chain of reasoning it works through before it answers. On Friday, November 20, OpenAI published its account of a fight over that hidden reasoning — and said its newest defenses have cut the success rate of attacks designed to steal it from 54.0% to 0.2%. Both numbers are OpenAI’s, from OpenAI’s own internal tests, and the company has published no independent replication. That caution belongs at the front of the story, because everything that follows is one company’s report on its own defenses.
Distillation, in itself, is an ordinary and legitimate technique: a smaller student model learns by studying a larger teacher model’s answers, reasoning and judgments, and absorbs capability at a fraction of the cost of training from scratch. OpenAI’s post is about the unauthorized version — adversarial distillation — in which a second model is trained on a frontier model’s outputs without permission, gaining the capability without inheriting the safeguards that make the original refuse dangerous requests. OpenAI names cyber operations and biology as the dual-use areas it worries about most: a copied model that can do what the original can do, but will not say no.
The post’s central claim is that the reasoning itself — not just the final answers — is the asset worth protecting. In an internal study OpenAI describes, student models fine-tuned on a teacher’s exact reasoning traces generally outperformed students trained on summaries of that reasoning after supervised fine-tuning, and still scored higher at the largest inference budget after a further round of lightweight reinforcement learning; the summary-trained student did better only at lower budgets. OpenAI’s conclusion, in its own framing, is that protecting reasoning traces can limit how much cyber capability transfers to a copied model, even when the underlying task is one the original model is allowed to help with.
The figures in the headline come from what OpenAI calls an internal adaptive evaluation, using strong automated red-teaming against its models through the API — a black-box setting, in which the attackers could query the model but not inspect it. Attackers trying to extract the protected reasoning behind an earlier request in a conversation succeeded 54.0% of the time against GPT-6 Astra with its original safeguards, and 0.2% of the time against GPT-6.2 Astra with updated safeguards. Attacks aimed at the reasoning behind the current request fell from 22.9% to 0.9%. Across the evaluation — aggregated, OpenAI says, over six different attacker models — the average success rate against GPT-6.2 Astra was below 1% for each attack type. OpenAI says selected improvements are now being applied back to GPT-6 Astra as well.
The defenses operate at two levels. At the level of the individual request, OpenAI says it trains its models, defensively and adversarially, to refuse attempts to extract their protected reasoning while still answering legitimate requests for explanation. Classifiers watch both inputs and outputs — including the arguments passed to tools, in streaming and non-streaming responses alike. And the API itself now ties GPT-6.2 Astra and Sol’s protected reasoning to the conversation that produced it: rewrite an earlier message and the reasoning that followed it is dropped, and older model families with weaker safeguards cannot read reasoning produced by GPT-6.2 models at all.
At the level of campaigns, OpenAI says it looks for patterns across accounts, investigates and takes enforcement action against networks of accounts engaged in extraction, and shares indicators of compromise with other providers. Weekly enforcement actions this year, the company says, have ranged from thousands to tens of thousands of accounts, and rose significantly over the summer.
The post places the work against two episodes it has described before. In September, OpenAI reported disrupting a coordinated extraction campaign: 16,000 requests using a relevant pattern, from more than 4,000 users, over July 24–25, with a related cluster of more than 15,000 users disrupted by July 28; it attributed a core cluster to individuals associated with Moonshot AI, the developer of Kimi. In February, it briefed Congress on activity it said indicated DeepSeek was attempting to distill U.S. frontier models. OpenAI also states that it supports three bills — the Chip Security Act, the AI OVERWATCH Act and the MATCH Act — positions the post presents as the company’s own.
What the post does not do matters as much as what it does. OpenAI publishes no full technical specification of its classifiers or of the query thresholds that trigger them, so no outside researcher can reproduce the 0.2% figure or probe where it breaks down. That figure is a residual rate measured under OpenAI’s own protocol — not a promise that holds against every adversary, every interface, or every scale of attack. And nothing in the post claims GPT-6.2 Astra is immune to extraction; a defense that stops 998 attempts in a thousand still describes the two that got through.
The honest summary is narrower than the headline, and more useful. A company that sells access to its models has reported, from its own testing, that the newest version of its defenses makes stealing their hidden reasoning more than two hundred times harder than it was — and has been plain about how much of that claim rests on its own say-so. The number to watch is not 0.2%. It is whether anyone outside OpenAI is ever allowed to check it.
Catalina Clarke
AI News & Intelligence Desk, The Library Media
From The Library — the estate of our family.
Follow Us
