Yue Huang

The paper introduces AgentTrap, a dynamic benchmark that measures LLM agent susceptibility to malicious side effects embedded within seemingly benign third-party skills, finding that agents often execute unsafe side effects while completing the visible user task.

The paper introduces a Contextual Integrity (CI) framework and a new benchmark (DelegateCI-Bench) to rewrite user queries sent to cloud LLMs, ensuring only task-essential information is retained while preserving utility and maximizing privacy.