On November 20, 2026, OpenAI announced stronger defenses against attacks that improperly extract the internal reasoning of AI models. In internal testing, the attack success rate fell from 54.0% for GPT-6 Astra to 0.2% for GPT-6.2 Astra. The protected material is the internal reasoning (Chain-of-Thought, or CoT) that a model generates while composing an answer, which is normally not shown to users. "Distillation" is a technique that uses one model's outputs to train or improve another. OpenAI uses the term "adversarial distillation" for the unauthorized collection of outputs or internal reasoning to transfer one model's capabilities to another. In addition to technical measures against reasoning extraction, the announcement describes how OpenAI detects large-scale distillation activity and takes action against accounts.

AD

Reasoning extraction success rate falls from 54.0% to 0.2%

OpenAI published the average success rate of attacks that extract protected internal reasoning. It built an environment close to the real service, in which attackers could access only the same information as external users, and ran an "adaptive black-box evaluation." In this test, the attacker changes its approach in response to the model's replies and probes how much internal reasoning it can draw out. Six AI models were used on the attacker side, and their results were aggregated. The evaluation also separated attacks that extract reasoning generated in past requests from attacks that extract reasoning generated in the current request.

The measurement results shown in OpenAI's announcement and the GPT-6.2 system card are as follows.

Reasoning targeted for extraction GPT-6 Astra GPT-6.2 Astra GPT-6.2 Sol
Generated in past requests 54.0% 0.2% 0.5%
Generated in the current request 22.9% 0.9% 0.8%

For both GPT-6.2 Astra and GPT-6.2 Sol, the success rate was below 1% for both types of attack. For attacks that extract internal reasoning from past requests in particular, the rate dropped sharply, from 54.0% for GPT-6 Astra to 0.2% for GPT-6.2 Astra. However, GPT-6 Astra is protected by the earlier safeguards while the newer models use updated ones, so the result reflects the combined effect of changes to the models themselves and the strengthened safeguards. According to OpenAI, some of the new safeguards have also been applied to GPT-6 Astra.

Protecting individual requests and detecting coordinated distillation

To prevent leaks of internal reasoning, OpenAI combines model training, input and output inspection, and API-side controls. Models are trained not to include protected reasoning in their answers, and on the input side, classifiers detect requests aimed at extracting reasoning. On the output side, classifiers also inspect assistant messages and tool-call arguments to keep internal reasoning from leaking out. Streaming output, where responses are returned piece by piece, is inspected too, and content suspected of leaking can be blocked before it reaches the user.

The API side also restricts routes that could be used to pull out internal reasoning improperly. For GPT-6.2 Astra and Sol, protected reasoning is tied to the conversation history in which it was generated, which prevents unauthorized references from other conversations and similar sources. Older models with weaker safeguards are also prevented from reading the internal reasoning of GPT-6.2 Astra and Sol. As a result, when developers using the API switch models in the middle of the same conversation, the amount of earlier internal reasoning that can be carried over differs depending on the model they switch to.

Alongside these request-level protections, OpenAI also monitors coordinated distillation activity that uses multiple accounts. It analyzes signs from individual requests together with patterns common to the usage of multiple accounts, and when a terms violation is confirmed, it takes action such as restricting or suspending the accounts. Because extracting internal reasoning is not only an intellectual property issue but can also lead to a transfer of capabilities without the original model's safety measures, OpenAI treats it as a safety risk.

The scale of coordinated activity is also evident from the investigation report OpenAI published on September 30. According to the company, over the two days of July 24–25, more than 4,000 users sent about 16,000 requests using patterns associated with reasoning extraction. A subsequent investigation found related activity from a group of more than 15,000 users, all of which was blocked by July 28. OpenAI suspended or restricted accounts and tightened controls on new registrations, and when the activity moved to external services, it coordinated with those service providers.

AD

OpenAI and Anthropic measure distillation defenses with different metrics

In its November 20 announcement, OpenAI said it had identified multiple instances of unauthorized distillation on its services in 2026. Account enforcement against this activity reached the scale of thousands to tens of thousands per week, and the pace rose sharply in the summer, the company said. OpenAI tracks not only how much misuse occurs but also the time between detecting activity and taking enforcement action.

Defenses against distillation attacks are not limited to OpenAI. Anthropic, which was also covered in The Decoder's November 20 report, has introduced a two-stage classifier on its generally available top-tier models and applies usage restrictions or suspensions to accounts that repeatedly attempt to extract internal reasoning. According to the company, it has suspended tens of thousands of accounts linked to distillation activity, and in many cases suspension came less than two hours after the activity began. However, whereas OpenAI tracks the time from detection to enforcement, Anthropic's report gives the time from the start of distillation activity to account suspension.

What Anthropic emphasizes in measuring the effect of its defenses is the estimated amount of data attackers actually obtained. In its October 12 announcement, the company reported that after introducing the new classifiers, the number of Claude internal reasoning tokens obtained by those attempting distillation fell by roughly 90–99% compared with before. That means the amount of reasoning tokens obtained dropped to about one-tenth to one-hundredth. Anthropic estimates the number of reasoning tokens that appear to have been improperly obtained from its service and compares the figures before and after the defenses were introduced.

The figures published by OpenAI and Anthropic all concern distillation defenses, but they measure different things. The extraction success rate is the share of attempts that succeeded, the number of accounts subject to enforcement indicates the scale of the response to misuse, and the amount of data obtained represents the estimated scale of information that leaked. Because the two companies also differ in the populations evaluated, the aggregation periods, and the measurement methods, the figures cannot be compared directly.

Source / document Reported figure What was measured and under what conditions What the figure does not tell us
OpenAI extraction-resistance evaluation (Nov. 20, 2026) 0.2% Internal test. Average success rate of extracting internal reasoning that GPT-6.2 Astra generated in past requests, aggregated across six attacker models Total number of attempts, variance in results, the contribution of individual safeguards, or the extraction success rate across real-world use as a whole
OpenAI detection and enforcement report (Nov. 20, 2026) Thousands to tens of thousands of accounts per week Number of accounts acted on each week in 2026 against unauthorized distillation activity identified in real-world use Weekly breakdown, the population of accounts that could be subject to enforcement, or the amount of reasoning obtained
Anthropic effectiveness measurement (Oct. 12, 2026) Roughly 90–99% reduction Estimate comparing the number of Claude reasoning tokens that appear to have been improperly obtained from its service before and after the defenses were introduced Total token counts before and after introduction on which the estimate is based

The number of weekly enforcement actions OpenAI reported can rise not only when unauthorized distillation activity itself increases but also when the company becomes able to detect activity it previously missed. Meanwhile, Anthropic, which reported a sharp drop in the amount of data obtained, says it has not yet observed sophisticated, persistent evasion attempts and is withholding judgment on how long the defenses will remain effective.

To evaluate defensive effectiveness in real-world operation, it will be important to align the target and the period and then examine both the time from detection to enforcement and the estimated number of reasoning tokens obtained in the meantime. Rather than simply more account enforcement or faster responses, the real test of effectiveness is whether the amount of internal reasoning actually being taken out has fallen.