The paper proposes an autonomous red teaming framework combining LLMs and RL to generate sophisticated, multi-stage cyber attack campaigns, demonstrating its necessity for evaluating robust AI-enabled SOAR systems.