THE article discusses the development of AI security agents, focusing on the training methods used by Dreadnode, an AI offensive security startup. Researchers found that red team agents, which simulate attacks, are significantly more effective than blue team agents, which defend against those attacks. To address this imbalance, Dreadnode released two open-source tools: DreadGOAD, a training environment for Active Directory, and Ares, a red-blue team system for evaluating agent efficiency.
The study reveals that blue agents struggle with data management and reasoning under pressure, but with improvements in their operational processes, they can enhance their defensive capabilities. Performance improvements in blue agents have also led to reduced operational costs.