SkillVetBench: LLM-as-Judge for Multi-Dimensional Security Risk Evaluation in Open-Source LLM Agent Skills

SKILLVETBENCH has been introduced as a public leaderboard on Hugging Face that employs an LLM-as-Judge to evaluate the security of community-contributed skills for open-source LLM agents. It features the Skill Agentic Risk Score (SARS), a five-dimensional metric that integrates CVSS v4.0 for comprehensive risk assessment, demonstrating zero false negatives in testing against confirmed malicious skills and outperforming static baselines like SKILLSIEVE. This tool addresses significant vulnerabilities in instruction-layer security, particularly in detecting threats such as prompt injection and memory poisoning, which conventional tools struggle to identify, making it essential for practitioners focused on secure AI deployments.