Minim Enhances Agent Privacy with Local UI State Sanitization.

Autonomous agents powered by Large Language Models often rely on extensive user interface (UI) state data to perform actions in digital environments. A common practice involves transmitting the entire UI state to remote servers, which poses a significant privacy risk by potentially exposing sensitive information like authentication codes or private notifications, even if irrelevant to the task. MINIM is proposed as a solution to this privacy concern. It functions as a trusted local broker on the client device, sanitizing UI observations before they leave the device. This system is based on Contextual Integrity principles and assigns each UI element a dual score: an inherent sensitivity score and a task-conditioned necessity score. These scores guide a ternary disclosure policy, which either keeps essential elements, abstracts sensitive attributes, or removes irrelevant content. By optimizing for a Contextual Integrity-aware objective, MINIM effectively prunes sensitive, non-essential data while preserving the critical semantic context and interactive elements needed for reliable agent operation. Experiments show it substantially reduces sensitive data leakage.