Large Language Models (LLMs)-based multi-agent systems (MAS) are vulnerable to the rapid propagation of toxicity, hallucinations, and adversarial attacks. Traditional defenses, which rely on static network topologies or binary “blocklists”, either fail to contain influential “super-spreaders” or unnecessarily fragment the network after minor errors. To address this, we propose EMAS, an Epidemiological Multi-Agent System framework with probabilistic connections. EMAS is a self-regulating network topology in which agent interactions are governed by probabilistic permeability rather than fixed connections. Drawing inspiration from epidemiological models, our framework dynamically adjusts communication bandwidth based on the interplay between a sender’s risk profile and a receiver’s robustness. This approach isolates malicious agents without manual intervention while preserving the flow of healthy information.