AIToday
AI Business & IndustryOpenAI BlogPublished: Nov 21, 2026, 06:00 JST

OpenAI: GPT-6.2 Astra cuts reasoning extraction attacks to 0.2%

Send on LINE
OpenAI: GPT-6.2 Astra cuts reasoning extraction attacks to 0.2%

3 Key Points

  1. What happened

    OpenAI said internal adaptive red-teaming showed the average attack success rate for recovering protected reasoning fell from 54.0% against GPT-6 Astra with its original safeguards to 0.2% against GPT-6.2 Astra with updated safeguards.

  2. Why it matters

    OpenAI frames adversarial distillation (using a frontier model's outputs to train another model without its safeguards) as a safety and national security risk, so stronger protection against reasoning-trace extraction may make it harder to copy advanced capabilities, though OpenAI plans to share more about recent campaigns.

WHO IT HITSModel developers and platform operators offering frontier AI models, plus their security and policy teams, may need to adopt similar request-level and campaign-level defenses against unauthorized capability extraction.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Context & Analysis

OpenAI's disclosure links technical safeguards to a broader policy argument. The company says adversarial distillation lets a second model learn advanced frontier capabilities without inheriting the safeguards that govern how the original model can use them, and it describes the concern as especially clear in dual-use domains like cybersecurity and biology. Its approach combines adversarial testing, model training, classifiers, and API-level protections with investigations into coordinated extraction attempts.

The company says it has strengthened protections for protected reasoning in GPT-6.2 Astra and GPT-6.2 Sol, tying a model's protected reasoning to the conversation history that produced it and blocking older model families with weaker safeguards from reading it. It also says an internal adaptive evaluation using strong automated red-teaming found the average attack success rate for recovering protected reasoning fell from 54.0% to 0.2%.

OpenAI connects the technical work to national security, citing a February briefing to Congress on activity indicating that DeepSeek was attempting to distill U.S. frontier models and a September report attributing a core part of a disrupted extraction campaign to individuals associated with Moonshot AI, the developer of Kimi. It says it supports legislation and export controls, and it recommends steps other developers can take, such as keeping successful extraction attack logs as regression tests and combining request-level mitigations with campaign-level enforcement.

FAQ
What is adversarial distillation?
It is an attempt to use a frontier model's outputs without authorization to train another model with similar capabilities, which is then deployed without comparable safeguards.
What are protected reasoning traces and why does OpenAI protect them?
They are the intermediate steps a model generates internally while working toward an answer, known as its chain-of-thought. OpenAI says these traces can contain information the final answer leaves out, making them especially valuable training data for another model.
How does OpenAI enforce against distillation campaigns?
OpenAI says its weekly enforcement actions against these campaigns have ranged in scale from thousands to tens of thousands of accounts, with the rate increasing significantly this past summer.
What legislation does OpenAI support on this issue?
OpenAI said it supports legislation giving the U.S. government more tools to identify and sanction foreign actors, and it told Congress it supports the Chip Security Act, the AI OVERWATCH Act, and the MATCH Act.
Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.

Questions and answers are published on this page.

Related Articles

Next articleOxide’s Cantrill: $445 million bet that AI compute moves on-prem