CDI: Contextualized Defense Instructing Achieves 94.2% Privacy Preservation for LLM Agents at 80.6% Helpfulness

arXiv·medium signalXBlueskyLinkedInCopy link

A new arXiv paper introduces Contextualized Defense Instructing (CDI), a privacy defense for LLM agents that inserts a step-specific instructor model to generate privacy guidance during execution rather than blocking outputs post-hoc. CDI achieves 94.2% privacy preservation while retaining 80.6% task helpfulness — a significant improvement over prior approaches that sacrifice utility for safety. The method is drop-in compatible with existing agent frameworks, making it immediately actionable for builders deploying agents on user data.