"Nuclear Deployed!": Analyzing Catastrophic Risks in Decision-making of Autonomous LLM Agents