Dive into the Agent Matrix: A Realistic Evaluation of Self-Replication Risk in LLM Agents