Reproduce “SkillScope: Toward Fine-Grained Least-Privilege Enforcement for Agent Skills” 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
Jiangrong Wu, Yuhong Nan, Yixi Lin, Huaijin Wang, Yuming Xiao, Shuai Wang, Zibin Zheng
SkillScope checks whether agent Skill actions are authorized and necessary for a user’s task, then gates confirmed over-privileged actions by task. It tests original and replay executions. On 200 annotated Skills, it achieves 94.53% skill-level F1; on 249 legitimate tasks, it completes all 249 after constraining, versus 144 with direct removal. These results apply to the paper’s evaluations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.