SafeMCP: Proactive Power Regulation for LLM Agent Defense via Environment-Anchored Lookahead Reasoning

SafeMCP is a server-side defense plugin for LLM agents using MCP that performs lookahead reasoning through an internal world model, constraining dangerous power expansion with two layers of defense: proactive tool filtering and immediate intervention. Its training uses a three-stage pipeline: environment dynamics anchoring, safety policy initialization, and reinforcement learning with dual verifiable rewards. Experiments on PowerSeeking Bench, ToolEmu and AgentHarm show that SafeMCP achieves a safety equilibrium, effectively mitigating risk while maintaining agent utility.