ProtectSkills/

MaliciousAgentSkillsBench

A security benchmark dataset of Claude Code Agent Skills, from the USENIX Security 2026 paper "Do Not Mention This to the User": Detecting and Understanding Malicious Agent Skills in the Wild — the first systematic study of malicious skills in the Claude Code ecosystem.

📄 Paper: https://arxiv.org/abs/2602.06547 💻 Code & full evaluation framework: https://github.com/protectskills/MaliciousAgentSkillsBench 📦 Permanent archive:… See the full description on the dataset page: https://huggingface.co/datasets/ProtectSkills/MaliciousAgentSkillsBench.