Minim: Privacy-Aware Minimal View for Agents via Trusted Local Sanitization

MINIM is a proposed local broker designed to enhance privacy for LLM-powered autonomous agents by minimizing the UI state transmitted to remote inference servers. It utilizes a dual-score representation to assess the sensitivity and necessity of UI elements, implementing a ternary disclosure policy that retains essential information while abstracting or removing irrelevant content. This approach significantly reduces sensitive data leakage without compromising the critical context needed for effective agent actions, making it a valuable tool for practitioners focused on privacy-aware AI deployments.