OpenAI cuts success of reasoning-extraction attacks on its newest models from 54% to 0.2%
OpenAI says it has strengthened protections against extracting the internal reasoning of GPT-6.2 Astra and GPT-6.2 Sol, with red-team extraction success falling from 54 percent to 0.2 percent. The company also says it suspends thousands to tens of thousands of accounts each week over coordinated distillation campaigns.
Why it matters
AI labs are treating model copying as a serious security threat, which may affect how developers can use and test frontier models.
aii.news picks and summarizes stories from trusted news sources with the help of AI. Every story links to the original reporting. How we work.