OpenAI Cut the Success Rate of Reasoning-Theft Attacks from 54% to 0.2%
OpenAI says its defense against reasoning theft cut the success rate of such attacks from 54% to 0.2%. The figures come from the company's internal tests. The idea behind the attack is simple: flood a model with queries, collect its detailed answers, and train a cheaper copy on them. This is called distillation. If the defense holds up in the real world, closed APIs will become stricter about what they reveal about a model's chain of thought.
AI-processed from X: @soluunj; edited by Hamidun News
The most valuable part of a modern language model is not the interface or the API, but how it reasons. Those long chains of steps are what let it solve hard problems in math, code, and science. So it is no surprise that a real race has started around them: some labs try to hide them, others try to copy them carefully. OpenAI says its new defense sharply reduced the odds of such copying. In its tests, the share of successful attacks aimed at stealing reasoning fell from 54% to 0.2%.
What is being stolen, and why
The technique in question is distillation. Imagine a strong tutor and a beginner competitor. The competitor asks the tutor thousands of questions, records not only the answers but the detailed explanations, and then trains their own student on those notes. In the end, the cheaper model picks up a problem-solving style that another company spent huge sums developing.
In its legitimate form, distillation is common practice: companies compress their own large models into smaller ones. The problem arises when someone does this to another company's model through a public API without permission. Until recently, model weights were considered the main asset. Now the reasoning itself, which can be extracted through ordinary queries, has become just as valuable.
What is known about OpenAI's defense
The announcement says little, but the main points are clear. OpenAI says it has deployed measures against reasoning theft and tested them on simulated attacks. Without the defense, attackers reached their goal in 54% of cases, roughly every second attempt. With the defense, that figure fell to 0.2%, or one successful attempt in five hundred. The company describes this as part of a broader effort against distillation.
The original announcement gives no further details. It does not disclose how the defense works: whether it hides part of the reasoning, adds noise, restricts suspicious query patterns, or does something else. It is unknown which models and attack scenarios were used in testing, how "success" of a theft was defined, or how much the defense affects answer quality for ordinary users. The figures come from the company itself, and there is no independent confirmation yet.
What this changes for developers and product managers
First, do not treat 0.2% as a guarantee. The test checks known attack methods, and real adversaries adapt. Such defenses usually work like a turnstile: they do not make theft impossible, but they raise its cost and slow it down.
Second, the defense may carry a cost for legitimate users. If a provider hides or shortens reasoning, it becomes harder to debug an agent, understand why it made a mistake, and check intermediate steps. Teams building products on closed APIs should plan for the possibility that access to reasoning details will narrow, and build quality control around verifying the final result rather than parsing someone else's "thoughts."
Third, if you offer a model or agent through your own API, you face the same risk. Basic measures are available now: limits on request volume per account, monitoring for unusually systematic access, terms of service that prohibit using outputs to train competing models, and withholding unnecessary internal information from responses.
Finally, the news shows where the industry is heading. Closed labs are increasingly protecting not just their weights but also the model behavior that can be read from the outside. This matters for open models too: the tighter access to the "thoughts" of closed systems becomes, the more disputes there will be about where the line lies between training on open data and copying someone else's work. For now, the main takeaway is modest: the claim is interesting and the scale is impressive, but it will have to be tested over time and by independent researchers.
- Nov 22, 2026Microsoft's Agent Hand-Wrote Neural Network Weights and Hit 100% Accuracy on 13 Tasks
- Nov 22, 2026LangChain, a Side Project Turned $3 Billion Startup in Four Years, Raises $300 Million
- Nov 22, 2026Android work profile closes real-IP leak that Happ and v2rayNG don't catch
- Nov 22, 2026Anthropic Goes Public on Nasdaq: Shares Rise from $83 to $111, but Can Safety Hold Up?
Key Questions Answered
What is being stolen, and why?
The technique in question is distillation. Imagine a strong tutor and a beginner competitor. The competitor asks the tutor thousands of questions, records not only the answers but the detailed explanations, and then trains their own student on those notes. In the end, the cheaper model picks up a problem-solving style that another company spent huge sums developing. In its legitimate form, distillation is common practice: companies compress their own large models into smaller ones.
What is known about OpenAI's defense?
The announcement says little, but the main points are clear. OpenAI says it has deployed measures against reasoning theft and tested them on simulated attacks. Without the defense, attackers reached their goal in 54% of cases, roughly every second attempt. With the defense, that figure fell to 0.2%, or one successful attempt in five hundred. The company describes this as part of a broader effort against distillation. The original announcement gives no further details.
What this changes for developers and product managers?
First, do not treat 0.2% as a guarantee. The test checks known attack methods, and real adversaries adapt. Such defenses usually work like a turnstile: they do not make theft impossible, but they raise its cost and slow it down. Second, the defense may carry a cost for legitimate users. If a provider hides or shortens reasoning, it becomes harder to debug an agent, understand why it made a mistake, and check intermediate steps.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.