Reproduce “When Lower Privileges Suffice: Investigating Over-Privileged Tool Selection in LLM Agents” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Kaiyue Yang, Yuyan Bu, Jingwei Yi, Yuchi Wang, Biyu Zhou, Juntao Dai, Songlin Hu, Yaodong Yang

Research explored "over-privileged tool selection" in large language model agents, where agents unnecessarily choose higher-privilege tools even when lower-privilege alternatives are sufficient, finding it widespread (over 30% in most models) and exacerbated by transient tool failures. A novel privilege-aware post-training approach significantly reduced this behavior, for example, lowering Qwen3-8B's over-privileged tool use rate from 64.9% to 27.02%.

Key insights

  • Over-privileged tool selection is a prevalent behavioral issue among diverse mainstream LLM agents, with a tendency to unnecessarily use high-privilege tools.
  • Transient execution failures with lower-privilege tools amplify agents' propensity for privilege escalation, as agents often fail to persistently explore conservative options.
  • Existing general safety alignment techniques, designed to prevent harmful outputs, do not inherently instill a preference for minimal privilege among authorized actions.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.