✉ The Friday AI Brief: the week's 5 best AI stories, tools & comparisons — in your inbox every Friday morning.

OpenAI logo

OpenAI Says It Cut Reasoning-Theft Attacks From 54% to 0.2% — Here’s How

OpenAI says its newest safeguards cut attackers’ success at stealing its models’ hidden reasoning from 54.0% to 0.2%, in a detailed account of how it detects and disrupts “adversarial distillation” published November 20. Why it matters: distillation is how a rival can copy a frontier model’s abilities without copying its safety rules — and OpenAI is now treating the defense as a national-security issue, not just a terms-of-service one.

November 20, 2026

What happened

Distillation is a standard technique in which one model learns from another model’s outputs. OpenAI’s concern is the adversarial version: an unauthorized actor systematically collects a frontier model’s outputs — above all its “protected reasoning,” the internal chain-of-thought steps a model generates on the way to an answer — and uses them to train a competing model that inherits the capability but not the safeguards. Those traces, OpenAI notes, can contain information the final answer leaves out, which makes them especially valuable training data.

In an internal study, OpenAI found that a student model trained on a teacher’s exact reasoning traces generally outperformed one trained only on the summaries users normally see, including on cybersecurity benchmarks after further training. Its conclusion: protecting the traces can limit how much risky cyber capability transfers, even when the original task was an allowed one.

The defenses, and the numbers

OpenAI describes two layers. At the request level, GPT-6.2 Astra and GPT-6.2 Sol are adversarially trained to refuse reasoning-extraction attempts while still answering legitimate questions; input classifiers flag extraction attempts and output classifiers scan both replies and tool-call arguments for leaked reasoning. At the API level, a model’s protected reasoning is now tied to the conversation that produced it — rewrite an earlier message and the reasoning that followed is dropped — and older model families with weaker safeguards cannot read the new models’ reasoning at all.

At the campaign level, OpenAI looks for coordinated patterns across accounts, enforces against networks it identifies, and shares indicators of compromise with other providers. It says its weekly enforcement actions against distillation campaigns this year have ranged from thousands to tens of thousands of accounts, with the rate increasing significantly this past summer.

The headline figures come from OpenAI’s own adaptive black-box red-teaming, in which automated attackers adjusted their approach as they went: extracting protected reasoning from an earlier request succeeded 54.0% of the time against GPT-6 Astra with its original safeguards, and 0.2% against GPT-6.2 Astra with the updated ones; for the current request, the rate fell from 22.9% to 0.9%. These are OpenAI’s internal measurements, not independently verified results. OpenAI says it is now applying selected improvements back to GPT-6 Astra as well.

Why it matters

This is the defensive sequel to a story that has been building all year: in February OpenAI briefed Congress on activity indicating DeepSeek was attempting to distill U.S. frontier models, and in September it disrupted a reasoning-extraction campaign it attributed in part to individuals associated with Moonshot AI, the maker of Kimi. OpenAI says it supports legislation giving the U.S. government more tools against foreign extraction, and backs the Chip Security Act, the AI OVERWATCH Act and the MATCH Act. The practical takeaway for developers is that API scrutiny is tightening: repeated requests aimed at hidden reasoning, especially spread across accounts, are exactly the pattern OpenAI says it now hunts. Everything AI, tested — and, increasingly, everything AI, guarded.

Frequently asked questions

What is adversarial distillation?

It is the unauthorized use of a model’s outputs to train another model. Unlike ordinary distillation done with permission, the adversarial version aims to reproduce a frontier model’s capabilities without its safeguards, refusals and monitoring.

What is “protected reasoning”?

The internal chain-of-thought a model generates while working toward an answer. OpenAI argues these traces are richer training material than final answers or summaries, so extracting them is the highest-value target for distillers.

How effective are the new defenses?

By OpenAI’s own internal red-team measurement, average attack success fell from 54.0% to 0.2% for reasoning from an earlier request, and from 22.9% to 0.9% for the current request, comparing GPT-6 Astra’s original safeguards with GPT-6.2 Astra’s updated ones. The figures have not been independently verified.

Sources: OpenAI (“Combatting adversarial distillation,” November 20, 2026)

Leave a Comment

Your email address will not be published. Required fields are marked *

Get the 5 best AI tools every week

Top AI news, tools, and prompts — one short email. Free, unsubscribe anytime.

Run a newsletter of your own? Monetize and grow it with SparkLoop →

Scroll to Top