
OpenAI tightens safeguards against extracting reasoning from GPT-6.2 models
OpenAI says attacks that extract protected reasoning from prior requests fell from 54.0 percent on GPT-6 Astra to 0.2 percent on GPT-6.2 Astra, while attacks on active requests fell from 22.9 to 0.9 percent.
OpenAI has sharply tightened safeguards against extracting internal reasoning from GPT-6.2 Astra and GPT-6.2 Sol. In internal automated tests, attacks pulling protected reasoning from prior requests dropped from a 54.0 percent success rate on GPT-6 Astra to 0.2 percent on GPT-6.2 Astra, while attacks on active requests fell from 22.9 to 0.9 percent. The company is now bringing selected improvements to GPT-6 Astra as well.
What changed in the tests
The reported attack success rates come from adaptive black-box red teaming aggregated across six attacker models. For prior request extraction, the rate fell from 54.0 percent to 0.2 percent. For current request extraction, it fell from 22.9 percent to 0.9 percent. A lower rate is better.
How the defenses work
For individual requests, the systems combine targeted training, input and output filters for messages and tool calls, and API blocks that stop older models from reading protected reasoning. Against coordinated campaigns, OpenAI bans thousands to tens of thousands of accounts weekly.
Industry context
Anthropic is also fighting systematic distillation attacks and bans tens of thousands of accounts to curb theft of internal reasoning, reportedly succeeding in 90 to 99 percent of cases. Anthropic has also called for a coordinated response from industry and policymakers. Earlier coverage from Feb 23, 2026 said Anthropic accuses Deepseek, Moonshot, and MiniMax of stealing Claude's AI data through 16 million queries.
Sources: The Decoder
SiTech — AI-powered web development
We build fast, modern websites and bring AI into real business workflows. Have a project or a question? We'd love to help.