Contextualized Privacy Defense for LLM Agents

Researchers propose Contextualized Defense Instructing (CDI), a new privacy defense paradigm for LLM agents that uses reinforcement learning to generate context-aware privacy guidance during execution. The approach achieves 94.2% privacy preservation while maintaining 80.6% helpfulness, outperforming static defense methods.