Reproduce “SkillGuard: A Permission-Centric Framework for Agent Skill Security” 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
Shidong Pan, Xiaoyu Sun, Tianyi Zhang, Dianshu Liao, Kaiwen Yang, Zhenchang Xing
SKILLGUARD is a permission-centric framework designed to secure LLM agent skill ecosystems by mediating both skill-based context influence and action side effects. It demonstrated a reduction in attack success rate (ASR) by 14.7-18.7 percentage points on the SkillInject benchmark, while largely preserving agent task utility.
Key insights
- Agent skill security requires a "dual-plane governance" model, which must mediate both a skill's subtle influence on an agent's reasoning context and its overt action-oriented side effects.
- A canonical, expressive permission taxonomy, analogous to traditional OS permissions, can effectively categorize and govern diverse resource-targeted actions across heterogeneous agent skills.
- Integrating proactive security checks at critical agent lifecycle hooks enables comprehensive policy enforcement that covers not only explicit tool calls but also inferred actions within complex commands (e.g., shell scripts).
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.