Reproduce “SkillSieve: A Hierarchical Triage Framework for Detecting Malicious AI 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
Yinghan Hou, Zongyou Yang
SkillSieve screens agent-skill packages before installation using cheap static checks, then applies contextual analysis and a three-model jury to packages that remain high-risk. On a human-reviewed benchmark of 390 skills, its full pipeline achieved F1 0.929, with precision 0.912 and recall 0.945. These results apply to that benchmark, not the 49,592-skill archive.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.