Privacy in the Age of Agents with Niloofar Mireshghallah

Niloofar Mireshghallah discusses privacy risks from frontier language models and agents. She argues that verbatim memorisation is no longer the main concern; aggregation attacks that assemble personal details from scattered sources, inference attacks such as precise geolocation from images, and leakage between inputs and outputs in agentic workflows matter more. The conversation also covers whether reasoning models rely on parametric knowledge or context, failures of models in non-English languages, and research directions industry neglects, including privacy-preserving collaboration between small local models and large commercial ones.