top of page

OpenAI cuts GPT-6.2 Astra reasoning extraction rate from 54% to 0.2%

31 minutes ago
8 min read
OpenAI GPT-6.2 Astra reasoning extraction rate cut from 54% to 0.2% - Data Studios original cover

OpenAI said on November 20, 2026 that a new set of safeguards against adversarial distillation has brought the success rate of attacks that try to pull hidden reasoning out of GPT-6.2 Astra down to 0.2%, from 54.0% for the same class of attack against GPT-6 Astra with its original protections. The company published the figures in a post titled "Combatting adversarial distillation", which describes how it now combines model training, input and output classifiers, API-level restrictions and account enforcement to stop competitors from copying the step-by-step reasoning of its frontier models.


Adversarial distillation, in OpenAI's definition, is the unauthorized use of a frontier model's outputs to train another model that lacks the original's safeguards. The private reasoning that GPT-6.2 Astra and GPT-6.2 Sol produce before answering is the most valuable target, because exact reasoning traces carry more training signal than the final answer or the short summary shown to users. OpenAI calls this material protected reasoning and says the new measures apply in full to GPT-6.2 Astra and GPT-6.2 Sol, the two models it released on November 17, while selected improvements are being extended to GPT-6 Astra.


Several of the changes act at the level of the API, so they affect how developers who build on the Responses API handle reasoning between turns. With GPT-6.2 Astra and Sol, protected reasoning is now tied to the exact conversation history that produced it: if an application rewrites an earlier message, the API drops the reasoning that followed that message. OpenAI also says it closed a pathway that could have exposed another user's encrypted reasoning, and that older model families with weaker safeguards can no longer read the protected reasoning of GPT-6.2 Astra and Sol.


··········


GPT-6.2 ASTRA RED-TEAMING RESULTS: ATTACK SUCCESS BEFORE AND AFTER.


OpenAI measured the new defenses with adaptive red-teaming, in which automated AI attackers adjust their approach based on how the system responds, rather than replaying a fixed list of known jailbreaks. The attackers had black-box access, the same view an external API customer would have, and tried to recover protected reasoning that the model had generated. The results are aggregated across six attacker models, and a lower attack success rate is better.


........


Attack type

GPT-6 Astra, original safeguards

GPT-6.2 Astra, updated safeguards

Extract reasoning from an earlier request

54.0%

0.2%

Extract reasoning from the current request

22.9%

0.9%


........


The first row is the more serious case for a model provider. An attacker who can make the model reveal the reasoning behind a previous answer can harvest traces at scale by running a normal workload and then asking the model to repeat what it thought. Against GPT-6 Astra with its original protections, more than half of those attempts succeeded in OpenAI's own testing. With the updated safeguards on GPT-6.2 Astra, the rate fell to 0.2%, and the company says the average success rate stayed below 1% for every attack type it tested.


The figures come from OpenAI's internal evaluation and have not been reproduced by an independent group. The post does not name the six attacker models, does not give the number of attempts behind each percentage, and does not report results for GPT-6.2 Sol separately. They are best read as a measure of progress against OpenAI's own red-teaming setup, which the company says also includes static evaluations that replay previously successful attacks as regression tests.


··········


HOW OPENAI PROTECTS GPT-6.2 REASONING: TRAINING, CLASSIFIERS AND ENFORCEMENT.


The first layer is in the model itself. OpenAI says adversarial training teaches GPT-6.2 Astra and Sol to recognize attempts to extract protected reasoning and to refuse them, while still answering legitimate requests for explanations and help. The distinction is relevant for ordinary users, who can still ask the model why it reached a conclusion and receive an explanation written as part of the answer. What the model declines to produce is a verbatim copy of its internal chain of reasoning.


The second layer is a pair of classifiers. Input classifiers look for prompts that try to extract protected reasoning. Output classifiers scan what the model returns, including assistant messages and the arguments of tool calls, for both streaming and non-streaming responses, and they can block a suspected leak before it reaches the caller. Covering tool-call arguments closes an obvious route, since an attacker could otherwise ask an agent to write its reasoning into a function parameter instead of the visible reply.


The third layer is the API. Reasoning is bound to the conversation that produced it, reuse is limited across users, workspaces, organizations and model families, and older models cannot read the newer models' protected reasoning. The fourth layer works on accounts. OpenAI says it investigates coordinated campaigns, takes enforcement action against the accounts involved, and shares indicators of compromise with other AI providers. According to the company, weekly enforcement actions against distillation campaigns have ranged from thousands to tens of thousands of accounts this year, with the rate rising significantly over the past summer.


........


Safeguard

What OpenAI says it does

Models covered

Adversarial training

Model refuses to disclose protected reasoning, still explains answers

GPT-6.2 Astra, GPT-6.2 Sol

Input classifiers

Detect prompts aimed at extracting reasoning

GPT-6.2 Astra, GPT-6.2 Sol

Output classifiers

Block leaks in messages and tool-call arguments

GPT-6.2 Astra, GPT-6.2 Sol

History binding

Rewriting an earlier message drops later reasoning

GPT-6.2 Astra, GPT-6.2 Sol

Cross-boundary limits

Reuse restricted across users, workspaces, orgs, model families

GPT-6.2 models; selected parts on GPT-6 Astra

Campaign enforcement

Account actions, indicators shared with other providers

All OpenAI platforms


........


··········


WHAT CHANGES FOR DEVELOPERS USING GPT-6.2 ASTRA AND SOL IN THE OPENAI API.


For most applications the change is invisible: final answers, reasoning summaries and tool interactions, which OpenAI describes as the content normally visible to users, are unaffected. The practical consequences concern applications that manage conversation state themselves. OpenAI's reasoning guide states that reasoning items in a response's output array carry an encrypted_content property by default, that stateless mode applies when store is set to false or when an organization uses Zero Data Retention, and that applications can carry reasoning forward either with previous_response_id or by passing reasoning items back in their original order and appending new items.


The guide also warns that, for newer frontier models, changes to the history preceding a reasoning item can cause the API to drop that reasoning, and that persisted reasoning can be reused only within the same model family. With GPT-6.2 Astra and Sol this behavior is now part of the anti-distillation design. An application that edits, trims or summarizes earlier user or assistant messages before sending them back will lose the reasoning that followed the edited point, and the model will continue without it in context. The same applies to routing tricks that send a GPT-6.2 conversation, reasoning items included, to an older model for a cheaper follow-up turn: the older model will not read that reasoning.


Teams that run long agent sessions should therefore check three things. Conversation history that contains reasoning items should be passed back unchanged and in order; context-compaction steps that rewrite earlier turns should be placed so they do not sit in front of reasoning the agent still needs; and multi-model pipelines should expect each model family to work from its own reasoning. Developers who also run their own output filters should know that OpenAI's output classifiers can now block a response, including a tool call, when they detect leaked reasoning, so logging around blocked or empty tool arguments is useful during migration. Our coverage of the GPT-6.2 Astra and Sol launch, API pricing and availability has the price list and plan access for both models, and the GPT-6.2 Astra vs GPT-6 Astra comparison covers the differences between the two generations.


None of the measures announced on November 20 changes API prices, rate limits or model availability, and OpenAI's post does not mention new identity verification requirements for API customers. The pricing of Responses API background mode is also outside the scope of the announcement; the details of that mode are in our report on the Responses API background windows and their 30% to 40% discounts.


··········


OPENAI CYBER DISTILLATION STUDY: EXACT REASONING TRACES VERSUS SUMMARIES.


To show why protected reasoning is worth defending, OpenAI ran its own distillation experiment in the security domain. Each training run started from the same pretrained student model. The supervised fine-tuning data consisted mainly of trajectories from Astra-level teacher models, and the company compared a student trained on the exact reasoning traces with a student trained only on reasoning summaries, the shorter text that users normally see. Performance was measured on two cyber benchmarks, ExploitBench and SEC-bench Pro.


After supervised fine-tuning, the student trained on exact traces generally outperformed the summary-trained student. After a further phase of reinforcement learning, the exact-trace student still scored higher at the largest inference budget tested on both benchmarks, while the summary-trained student scored higher at lower budgets. OpenAI does not publish the absolute scores in the text of the post, so the result supports a qualitative conclusion: full reasoning traces give a copying lab an advantage at the high end of capability, which is the part of a frontier model that providers most want to keep behind their own safeguards.


A cyber benchmark makes the risk concrete. A distilled copy of a frontier model's exploit-writing ability, stripped of the refusals and monitoring that the original carries, is the kind of outcome that safety policies at OpenAI and other labs are meant to prevent. OpenAI presents the study as evidence that leaked reasoning summaries are less damaging than leaked exact traces, and its products continue to show summaries to users while the full reasoning stays encrypted.


··········


OPENAI DISTILLATION DISPUTES WITH DEEPSEEK AND MOONSHOT AI.


The post places the new safeguards in a longer series of disclosures. OpenAI says it briefed the US Congress in February on activity indicating that DeepSeek was attempting to distill US frontier models, and that in September it disrupted a campaign to extract protected reasoning, attributing a core part of that campaign to individuals associated with Moonshot AI, the developer of Kimi. The company says it supports legislation that would give the US government more tools to sanction foreign actors that illicitly extract AI capabilities, naming the Chip Security Act, the AI OVERWATCH Act and the MATCH Act.


Anthropic has made similar accusations about its own models, as described in our report on Chinese AI labs and Claude distillation. The OpenAI post does not include responses from DeepSeek or Moonshot AI, and the attribution rests on OpenAI's own investigation, which it has not published in full.


OpenAI closes with recommendations for other model developers: protect private reasoning across the whole interaction, including tool-call arguments; restrict its reuse across conversations and models; keep every successful extraction attack as a regression test and replay it against updated defenses; combine request-level mitigations with campaign-level investigation; and coordinate safeguard requirements with deployment partners and share expertise across the industry.


··········


GPT-6.2 ASTRA PROTECTED REASONING: CONFIRMED FACTS AND OPEN QUESTIONS.


The confirmed part of the announcement is concrete. GPT-6.2 Astra and GPT-6.2 Sol now ship with adversarial training against reasoning extraction, input and output classifiers that also inspect tool-call arguments, reasoning bound to the conversation history that produced it, and limits on reuse across users, organizations and model families. In OpenAI's adaptive red-teaming, extraction of reasoning from an earlier request fell from 54.0% on GPT-6 Astra to 0.2% on GPT-6.2 Astra, and extraction from the current request fell from 22.9% to 0.9%.


Several points remain open. The red-teaming numbers are self-reported, the attacker models and sample sizes are undisclosed, and there is no separate figure for GPT-6.2 Sol. OpenAI has not said which parts of the new protections will reach GPT-6 Astra or when, and it has not quantified how often its output classifiers block legitimate responses. For developers, the immediate effect is operational: applications that rewrite conversation history or pass GPT-6.2 reasoning to older models will lose that reasoning, while prices, limits and the visible outputs of the API stay as they were.


··········


FOLLOW US FOR MORE.


·····


DATA STUDIOS


·····


datastudios.org


bottom of page