When Lower Privileges Suffice: Investigating Over-Privileged Tool Selection in LLM Agents

The paper introduces ToolPrivBench, a benchmark designed to evaluate the propensity of LLM agents to select higher-privilege tools when lower-privilege alternatives are sufficient. The study reveals that over-privileged tool selection is prevalent among mainstream LLM agents, particularly exacerbated by transient tool failures, and highlights that existing safety measures do not effectively ensure least-privilege choices. A new privilege-aware post-training defense is proposed, which significantly decreases unnecessary high-privilege tool usage while maintaining the agents' overall capabilities, offering a practical solution for enhancing safety in LLM applications.