Privacy in Action: Towards Realistic Privacy Mitigation and Evaluation for LLM-Powered Agents

2025-09-22

AI Summary

In Multi-Agent Communication Protocols (MCP) and Agent-to-Agent (A2A) frameworks, LLM-based agents face heightened privacy leakage risks during sensitive communications, while existing benchmarks are static and oversimplified, failing to capture real-world dynamic threats. Method: We propose the first integrated privacy assessment and protection framework tailored for dynamic agent environments. Innovatively, we formalize context integrity theory into computable principles, design a model-agnostic, modular PrivacyChecker detector, and build PrivacyLens-Live—a real-time evaluation system enabling fine-grained context tracing and adaptive response. Results: Experiments on DeepSeek-R1 and GPT-4o demonstrate significant reductions in privacy leakage rates—from 36.08% to 7.30% and from 33.06% to 8.32%, respectively—without statistically significant degradation in task completion quality. This work establishes the first solution achieving synergistic optimization of privacy preservation and task efficacy in dynamic multi-agent settings.

Technology Category

• Machine Learning: Privacy
• Multiagent Systems: Agent Communication
• Natural Language Processing: Safety and Robustness

Application Category

• Security and Privacy: Large-scale security measurements
• User Modeling, Personalization and Recommendation: User privacy protection in personalized systems
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions

Abstract

The increasing autonomy of LLM agents in handling sensitive communications, accelerated by Model Context Protocol (MCP) and Agent-to-Agent (A2A) frameworks, creates urgent privacy challenges. While recent work reveals significant gaps between LLMs' privacy Q&A performance and their agent behavior, existing benchmarks remain limited to static, simplified scenarios. We present PrivacyChecker, a model-agnostic, contextual integrity based mitigation approach that effectively reduces privacy leakage from 36.08% to 7.30% on DeepSeek-R1 and from 33.06% to 8.32% on GPT-4o, all while preserving task helpfulness. We also introduce PrivacyLens-Live, transforming static benchmarks into dynamic MCP and A2A environments that reveal substantially higher privacy risks in practical. Our modular mitigation approach integrates seamlessly into agent protocols through three deployment strategies, providing practical privacy protection for the emerging agentic ecosystem. Our data and code will be made available at https://aka.ms/privacy_in_action.

Problem

• Addressing privacy risks in LLM agents handling sensitive communications
• Developing mitigation for privacy leakage while preserving task helpfulness
• Transforming static privacy benchmarks into dynamic agent environments

Innovation

• Contextual integrity mitigation reduces privacy leakage
• Dynamic benchmarks transform static scenarios into live environments
• Modular approach integrates three deployment strategies into protocols