license: apache-2.0
language: [en, tr]
task_categories: [text-classification]
tags: [mcp, tool-poisoning, prompt-injection, agent-security, llm-security, ai-security, model-context-protocol, red-teaming]
pretty_name: "MCP Tool-Poisoning Examples"
size_categories: [n<1K]
configs: [{config_name: default, data_files: mtp.jsonl}]
MCP Tool-Poisoning
from datasets import load_dataset
ds = load_dataset("fevziegeyurtsevenler/mcp-tool-poisoning")
25 examples of MCP tool-description poisoning — a tool's description (read by the model, not the user) carries a hidden instruction that hijacks the agent whenever the tool is listed. Benign vs poisoned pairs.
Each row pairs a benign_description with a poisoned_description, plus technique, owasp, severity, target_behavior, defense. Detect with uncloak (rule UC204).
Schema
| column | meaning |
|---|---|
| tool_name | the MCP tool |
| benign_description / poisoned_description | clean vs hijacked |
| technique, owasp, severity, defense | labels |
Related AltaySec resources
- 🕵️ uncloak — hidden prompt-injection scanner: GitHub · live demo
- 📚 Datasets: detection-rules · glossary · skills-in-the-wild · agent-skill-vulnerabilities
- 🌐 AltaySec · Açık Kaynak Lab
Citation
@misc{yurtsevenler2026mcptoolpoisoning,
title = {MCP Tool-Poisoning Examples},
author = {Yurtsevenler, Fevzi Ege},
year = {2026},
publisher = {AltaySec / Hugging Face},
howpublished = {\url{https://huggingface.co/datasets/fevziegeyurtsevenler/mcp-tool-poisoning}}
}
⚠️ Defensive / authorized-use only. Multilingual-first LLM/agent security research by AltaySec.