UC Santa Cruz

In recent years, there has been a focus on the potential of using artificial intelligence agents for network security. In this paper, we consider a scenario where network segmentation protocols require a multi-agent solution, along with an adversary that can evade traditional firewall solutions via email, known as the CAGE 4 challenge. While the challenge is focused on the use of artificial intelligence, we propose a solution that was manually crafted by analyzing the potential alerts provided by the environment with human eyes. The exceptionally good performance of our agent, especially when compared to other submissions to the challenge, leads us to raise questions about the efficacy of such challenges for evaluating the performance of artificial intelligence algorithms.