RESEARCH

OpenAI Cuts Reasoning-Extraction Attack Success From 54% to 0.2% in GPT-6.2 Astra

J James Whitfield Nov 21, 2026 2 min read
Engine Score 7/10 — Important

Security-relevant research

Editorial illustration for: OpenAI Cuts Reasoning-Extraction Attack Success From 54% to 0.2% in GPT-6.2 Astra
  • OpenAI says adaptive attacks that extracted protected reasoning from a prior request succeeded 54.0% of the time against GPT-6 Astra, versus 0.2% against GPT-6.2 Astra.
  • Attacks targeting the current, active request dropped from a 22.9% to a 0.9% success rate between the two model versions.
  • OpenAI combines targeted training, input and output filters, and API blocks that stop older models from reading protected reasoning, and bans thousands to tens of thousands of accounts weekly tied to coordinated extraction campaigns.
  • Anthropic has separately said it blocks 90 to 99% of systematic distillation attempts against Claude and has called for an industry-wide response.

What Happened

OpenAI said it sharply tightened safeguards against attackers trying to extract the internal reasoning of GPT-6.2 Astra and GPT-6.2 Sol, citing internal adaptive red-teaming results aggregated across six attacker models, according to The Decoder (Nov 20, 2026).

Why It Matters

The update extends an industry-wide push against model distillation that OpenAI, Anthropic, and Google coordinated on earlier this year, as labs try to stop competitors from cheaply copying frontier reasoning capability through extraction attacks.

Technical Details

In OpenAI’s tests, attacks extracting protected reasoning from a prior request fell from a 54.0% success rate against GPT-6 Astra’s original safeguards to 0.2% against GPT-6.2 Astra’s updated defenses. Attacks against the model’s current, active request dropped from 22.9% to 0.9% over the same comparison. The defenses combine targeted training, input and output filters on messages and tool calls, and API-level blocks that prevent older models from reading protected reasoning; OpenAI also bans thousands to tens of thousands of accounts weekly tied to coordinated extraction campaigns.

Who’s Affected

Developers building on OpenAI’s reasoning models benefit from reduced exposure of proprietary chain-of-thought data, while rival labs and distillation-focused startups lose a cheap channel for approximating frontier reasoning. OpenAI said it is also backporting select improvements to the original GPT-6 Astra.

What’s Next

Anthropic says it is separately blocking 90 to 99% of systematic distillation attempts against Claude, and has called for a coordinated industry and policy response to the broader extraction problem.

Share

Enjoyed this story?

Get articles like this delivered daily. The Engine Room — free AI intelligence newsletter.

Free · No spam · Unsubscribe anytime