Red Agents vs. Blue Agents: How to Make AI Better At Defense

Researchers at Dreadnode have developed open-source tools, DreadGOAD and Ares, to better evaluate the effectiveness of AI-powered security agents. They discovered that offensive (red team) agents consistently outperformed defensive (blue team) agents due to the difficulty in generating sufficient training data for defensive tasks and the inherent asymmetry of attack versus defense. To address this, they used red team agents to generate attack data, which was then applied to blue team agents, leading to noticeable improvements in their defensive capabilities. The team also tackled issues like context window limitations and agent disengagement during live attacks, ultimately demonstrating a path to more effective and cost-efficient AI-driven security.