CTDE large-scale multi-agent simulation, socialized LLM-MARL collaboration, agent privilege governance and large model adversarial red-team security research