Explore key findings from a 600-run benchmark study on AI agent architectures for offensive security, revealing optimal strategies for vulnerability detection, cost-efficiency, and deployment.