MaliciousAgentSkillsBench — Security skill for Claude Code

"Do Not Mention This to the User": Detecting and Understanding Malicious Agent Skills


This repository contains a comprehensive security benchmark dataset and evaluation framework for Claude Code Agent Skills. We collect __98,380 skills__ from two major platforms (skills.rest and skillsmp.com), including **157 malicious samples** identified through systematic security analysis.

Project Structure

MaliciousAgentSkillsBench/
├── data/                           # Benchmark datasets
│   ├── malicious_skills.csv        # 157 malicious skill samples (curated)
│   ├── skills_dataset.csv          # 98,380 total skills (157 malicious flagged)
├── code/                           # Security analysis framework
│   ├── analyzer/                   # AI-powered deep security analysis
│   ├── crawler/                    # Multi-platform data crawler
│   ├── executor/                   # Dynamic execution in Docker sandbox
│   ├── scanner/                    # Static rule-based security scanner
│   └── scripts/                    # analysis pipeline
│   └── ···                         # other files
└── README.md                       # This file

Disclaimer

__This repository contains examples of malicious agent skills for research purposes only. Reader discretion is recommended. Any misuse is strictly prohibited.__

The code and data in this repository are intended exclusively for:

• Academic research on AI agent security
• Developing defense mechanisms against malicious agent skills
• Evaluating the robustness of AI agent platforms

Data

Dataset Statistics

Source	Repos	Total Skills	Suspicious	Malicious
skills.rest	3,217	25,187	814	21
skillsmp.com	10,373	73,193	3,473	136
Total	13,590	98,380	4,287	157

Data Files

`malicious_skills.csv`

Curated dataset of **157 ve