OpenAI has significantly reduced the success rate of automated attacks designed to extract internal reasoning from its GPT-6.2 Astra and GPT-6.2 Sol models. Internal testing shows that attempts to pull protected reasoning from prior requests dropped from 54.0 per cent to 0.2 per cent on the updated version. Attacks on active requests fell from 22.9 per cent to 0.9 per cent. The company is now applying these selected improvements to GPT-6 Astra as well.
The defences operate at two levels. Individual requests face targeted training, input and output filters for messages and tool calls, and API blocks preventing older models from reading protected reasoning. Against coordinated campaigns, OpenAI bans thousands to tens of thousands of accounts weekly. This approach mirrors actions taken by Anthropic, which reports blocking 90 to 99 per cent of such distillation attempts while calling for industry and policymakers to respond in a coordinated way.




