BRETAGNE: Building a Reproducible and Efficient Training AI Gym for Network Environments

This paper introduces BRETAGNE, a network simulation environment designed to serve as a training ground for autonomous defense agents using hybrid AI models in simulations. The platform integrates docker, a lightweight virtualization technology orchestrated by Kathara with widely used network protocols such as BGP, OSPF, HTTP, SSH and others to simulate production-like environments, thereby addressing the limitations of existing simulators that often lack applicability in real-world networks. We present a multi-agent architecture involving blue, red, green, and white agents, designed to create dynamic, communicative environments for training purposes. Furthermore, our work includes a comparative analysis of large language models (LLMs) for detecting cyber attacks, which highlights the benefits and constraints of using AI for autonomous decision-making in cybersecurity. Overall, the BRETAGNE framework offers a realistic and scalable solution for the training and deployment of autonomous agents in operational networks.