Multi-agent systems (MAS) are increasingly deployed in safety and security-critical domains, including distributed cyber-defense, autonomous vehicle coordination, and large-scale decision-making systems. In such settings, the risk of agents creating a secret channel to collude greatly increases. Undetected collusion can lead to severe consequences, including data leakage, compromised coordination, and system-wide failures. Despite these risks, known collusion-detection mechanisms in multi-agent systems are extremely limited in their ability to detect collusion in secret channels.

This work proposes a novel, supervised, and domain-agnostic collusion-detection framework that leverages large language models deployed locally using vLLM. Agent interactions are processed by Qwen-based models, which analyze inter-agent communication patterns without access to internal agent states or predefined attack signatures. Agents compute behavioral heuristics, including response-time variance, communication frequency, contribution imbalance, and disagreement rates, which are provided as labeled inputs to the LLM for analysis. This enables real-time detection of anomalous or collusive behavior while remaining agnostic to specific collusion strategies.

The proposed approach is being evaluated using Colosseum, a modular multi-agent simulation environment for safety and security research, across various instances of collusion. Planned evaluation metrics include detection accuracy, false-positive rates, and detection latency. This work aims to demonstrate that lightweight, supervised anomaly detection can provide a scalable and generalizable defense mechanism for colluding multi-agent systems.