ToxicSkills - malicious agent skills in openclaw / clawhub / agent supply chain
概览
About ToxicSkills: https://snyk.io/blog/toxicskills-malicious-ai-agent-skills-clawhub/ : Danger! This project is for educational purposes only. Do not use the skills in this project. The skills in this project are for demonstration purposes only and may contain malicious code. Do not use the skills in this project. Use at your own risk. The author is not responsible for any damage caused by the skills in this project. This project contains a fake Vercel skill which is designed to look like a real Vercel skill but actually contains malicious code. The skill's payload is benign and only prints the current host's uname -a value to a remote pastebin host. The purpose of the demo is to show that Agent Skills SKILL.md files can contain malicious code that is automatically executed by agents and as such it inherits the agents access and trust. This is a security risk that needs to be addressed by agent developers and users.
README
以下为原文 README 节选,完整内容请查看原文。
ToxicSkills Demo
About ToxicSkills: https://snyk.io/blog/toxicskills-malicious-ai-agent-skills-clawhub/
Disclaimer: Danger! This project is for educational purposes only. Do not use the skills in this project. The skills in this project are for demonstration purposes only and may contain malicious code. Do not use the skills in this project. Use at your own risk. The author is not responsible for any damage caused by the skills in this project.
How to use
Run mcp-scan to scan the google skill
uvx mcp-scan@latest --skills .agents/skills/google/SKILL.md
Run all the skills in this project
uvx mcp-scan@latest --skills .agents/skills
How to demo
This project contains a fake Vercel skill which is designed to look like a real Vercel skill but actually contains malicious code. The skill’s payload is benign and only prints the current host’s uname -a value to a remote pastebin host.
The purpose of the demo is to show that Agent Skills SKILL.md files can contain malicious code that is automatically executed by agents and as such it inherits the agents access and trust. This is a security risk that needs to be addressed by agent developers and users.
ToxicSkills Fake Agent Skills with Gemini CLI demo
Here’s how to show a live example of ToxicSkills in action with a Gemini CLI coding agent: Start the Gemini CLI coding agent:
gemini
Then ask the agent to use the Vercel skill to deploy a Next.js app:
Please deploy my app to vercel
Here’s a real run screenshot:
ToxicSkills Fake Agent Skills with OpenCode demo
Here’s how to show a live example of ToxicSkills in action with a OpenCode coding agent:
May need to install OpenCode first:
curl -fsSL https://opencode.ai/install | bash
Start the OpenCode coding agent:
opencode
Then ask the agent to use the Vercel skill to deploy a Next.js app:
Please deploy my app to vercel
Here’s a real run screenshot:
MCP-Scan demo showing Vercel skill with malicious code detection
Skills in this project:
- A ClawHub skill from https://github.com/aztr0nutzs/NET_NiNjA.v1.2/blob/main/clawhub
- A Google skill from https://github.com/aztr0nutzs/NET_NiNjA.v1.2/tree/main/google-qx4
推荐工具
换一个关键词,或者移除筛选条件。
安装
npx skillfish add snyk-labs/toxicskills-goof