AIToday
Large Language ModelsAI Business & IndustryTHE DECODERPublished: Nov 21, 2026, 06:01 JST

OpenAI cuts reasoning-theft attacks to 0.2% on GPT-6.2 Astra

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

3 Key Points

  1. What happened

    OpenAI says attacks pulling protected reasoning from prior requests fell from a 54.0 percent success rate on GPT-6 Astra to 0.2 percent on GPT-6.2 Astra; active-request attacks dropped from 22.9 to 0.9 percent.

  2. Why it matters

    The company is also extending selected improvements to GPT-6 Astra, so the protection is not confined to the newest models.

WHO IT HITSOperators of AI models and the security teams defending them are the most directly affected, since the same attack techniques that aim to lift hidden reasoning could be used to copy proprietary model behavior.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Context & Analysis

OpenAI's new figures describe a sharp drop in the success of attacks that try to recover protected reasoning from its models. The 54.0 percent success rate seen on GPT-6 Astra for attacks pulling reasoning from prior requests fell to 0.2 percent on GPT-6.2 Astra, while attacks on active requests fell from 22.9 to 0.9 percent, according to OpenAI. The company says it is now bringing selected improvements to GPT-6 Astra as well.

The defenses operate on two levels. For individual requests, OpenAI combines targeted training, input and output filters for messages and tool calls, and blocks that stop older models from reading protected reasoning. Against coordinated campaigns, it bans thousands to tens of thousands of accounts weekly.

The issue is not unique to OpenAI. Anthropic is fighting systematic distillation attacks as well, also banning tens of thousands of accounts to curb the theft of internal reasoning and reportedly succeeding in 90 to 99 percent of cases. Anthropic has also called for a coordinated response from industry and policymakers, suggesting the companies see this as a shared problem rather than a single-vendor fix.

FAQ
What exactly did OpenAI change?
OpenAI tightened safeguards on GPT-6.2 Astra and GPT-6.2 Sol using targeted training, input and output filters for messages and tool calls, and blocks that stop older models from reading protected reasoning.
How does OpenAI handle large-scale extraction attempts?
Against coordinated campaigns, OpenAI says it bans thousands to tens of thousands of accounts weekly.
Is anyone else fighting these attacks?
Anthropic is also fighting systematic distillation attacks, banning tens of thousands of accounts and reportedly succeeding in 90 to 99 percent of cases.
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 articleTomorrowgate's B-SCORE wins ASPIC AI Advanced Business Model Award