OpenAI says GPT-6.2 cuts reasoning-extraction attacks to 0.2%
TL;DR
- OpenAI reports adaptive red-team success at extracting protected reasoning from an earlier request fell from 54.0% on GPT-6 Astra to 0.2% on GPT-6.2 Astra.
- The new safeguard ties protected reasoning to the conversation that produced it; if an earlier message is rewritten, the API drops the reasoning that followed.
- OpenAI names DeepSeek and individuals associated with Moonshot AI, with weekly enforcement actions ranging from thousands to tens of thousands of accounts.
OpenAI reports that adaptive red-team attacks extracting protected reasoning from an earlier request succeeded 54.0% of the time against GPT-6 Astra and 0.2% against GPT-6.2 Astra. For attacks on the current request, the rate fell from 22.9% to 0.9%. The figures appear in a company post aggregating results across six attacker models in a black-box setting.
The mechanism is structural. "With GPT-6.2 Astra and GPT-6.2 Sol, we tie a model's protected reasoning to the conversation history that produced it," the company writes. "As new messages are added, the model can build on its earlier reasoning. If an earlier message is rewritten, our API drops the reasoning that followed it." The practical effect is that a user who edits a prior prompt loses whatever hidden reasoning it produced.
OpenAI frames the problem geopolitically. "In February, we briefed Congress on activity indicating that DeepSeek was attempting to distill U.S. frontier models," the post says, and in September it reported disrupting "a campaign to extract our models' protected reasoning and attributed a core part of that activity to individuals associated with Moonshot AI, the developer of Kimi." Weekly enforcement actions have "ranged in scale from thousands to tens of thousands of accounts, with the rate increasing significantly this past summer."
The post publishes black-box numbers; the six attacker models are not named, and no per-domain breakdown is given.
This is the second GPT-6.2 Astra alert on our tracker in two days, following yesterday's MedChemBench release, part of a run of 427 OpenAI items in the last 90 days.
Shared on Bluesky by 1 AI expert
Originally reported by openai.com
Read the original article →Original headline: OpenAI Says New GPT-6.2 Defenses Cut Reasoning-Extraction Attack Success From 54% to 0.2%