Definition
This is a technique where an attacker compares a standard AI model to a 'reasoning' version trained from it, using the differences in how the reasoning model 'thinks' to extract hidden information the model was never supposed to reveal — up to 10 times more effectively than prior methods. It exploits the extra internal reasoning steps that make modern AI models smarter as a new source of leakage.
Why it matters
As companies rush to deploy more powerful 'reasoning' AI models, this shows that added intelligence can come with new, harder-to-anticipate data-leakage risks that standard security testing may miss.