Don't Trust Stubborn Neighbors: A Security Framework for Agentic Networks

Large Language Model (LLM)-based Multi-Agent Systems (MASs) deliberate through communication to solve complex tasks like web automation and itinerary planning. However, this interactive nature makes them vulnerable to malicious agents that can exploit communication channels to manipulate collective decisions. To model these deliberation dynamics, we propose a theoretical framework based on the Friedkin–Johnsen opinion dynamics model. Despite its simplicity, it closely captures empirical LLM-MAS behavior across various network topologies and attack settings. Our framework's tractability enables us to derive conditions under which stubborn, persuasive agents can trigger cascades that dominate opinion formation. Building on our theoretical analysis, we introduce a trust-adaptive defense that dynamically adjusts inter-agent trust to reduce adversarial influence without sacrificing cooperative performance. Extensive experiments validate both our modeling framework and the defense mechanism, demonstrating improved robustness against adversarial takeovers.