The first attack on “privacy-preserving” LLM-based insights systems

CSE Professor Emiliano De Cristofaro and his students devised the first attack on “privacy-preserving” LLM-based insights systems, i.e., platforms like Anthropic’s Clio that analyze real-world chats with AI assistants to extract aggregate insights about how people use them, showing their privacy claims don’t hold up in practice.

The paper is accepted to NeurIPS '26 (oral). There were 30709 valid paper submissions to the NeurIPS Main Track this year, of which the program committee accepted 7900 (25.7%) papers in total, with breakdown of 7496 as posters, 292 as spotlight and 112 as oral.