Yidan Hu

The paper proposes ADAM, a novel and highly effective privacy attack that systematically extracts sensitive data from LLM agent memory by adaptively querying the victim agent's memory based on data distribution and entropy.

PrivScope is a novel on-device governor that enforces task-scoped disclosure, ensuring sensitive information is abstracted to the least revealing form before being sent to a cloud language model, significantly reducing data leakage without compromising task utility.