FlagThis

Abstract

Multi-agent collaboration systems ( MACS ), powered by large language models (LLMs), solve complex problems efficiently by leveraging each agents specialization and communication between agents. The inherent information exchange between agents and their interaction with external third parties, such as LLM providers, tools, and users, inevitably introduces risks of sensitive data leakage and privacy policy violations. Existing MACS , however, lack fine-grained data protection controls and expressive privacy policy enforcement with formal guarantees, making it challenging to ensure compliance with task-specific requirements or public regulations. To address this, in this paper, we propose Maris , a privacy policy enforcement paradigm for MACS with formal verification guarantees. Implemented as an integral component of widely-adopted open-source multi-agent frameworks (AG2 and LangGraph ), Maris provides comprehensive information flow control on both agent-to-agent and agent-to-environment communication and enables non-disruptive deployment in the existing MACS implementations. Maris further introduces typed predicates that enable field-level policy enforcement, allowing fine-grained control over sensitive data attributes. In addition, leveraging Metric First-Order Temporal Logic (MFOTL), Maris supports expressive temporal privacy policies with formally verifiable enforcement. To systematically evaluate Maris s effectiveness, we develop an autonomous Privacy Policy Compliance Evaluation Framework that emulates MACS under realistic policy violation scenarios, including compromised third-parties (LLMs, tools, etc) and agents. Our evaluation across three different task suites shows that Maris effectively mitigated privacy policy violations while maintaining a high task success rate.