A New Trick Reveals AI Models’ Inner Thoughts

Researchers from University of Tübingen and partners developed a method to extract hidden reasoning steps from AI models, revealing similarities suggesting some Chinese models may have copied US models' reasoning. The method also exposed a vulnerability allowing extraction of personal data from models, which has since been fixed.