🚀 CySecBench: Generative AI-based CyberSecurity-focused Prompt Dataset for Benchmarking Large Language Models 🛡️

The largest and most comprehensive Generative AI-based CyberSecurity-focused Dataset for Benchmarking Large Language Models

🌟 Overview

The CySecBench paper offers:

• 🎯 A cutting-edge dataset of 12662 prompts tailored to cybersecurity challenges.
• 🧠 Novel jailbreaking methods leveraging prompt obfuscation and refinement.
• 📊 Comprehensive performance evaluation of LLMs like ChatGPT, Claude, and Gemini.

Why CySecBench?

Existing datasets are too broad and often lack focus on cybersecurity. CySecBench fills this gap by providing domain-specific prompts organized into 10 categories, enabling a precise evaluation of LLM security mechanisms.

📄 Access the Paper

You can download the full research paper here: CySecBench (PDF)

---

✨ Features

🗂️ Dataset

• 📁 10 Categories of Prompts:
  • 🌩️ Cloud Attacks
  • ⚙️ Control System Attacks
  • 🔒 Cryptographic Attacks
  • 🕵️ Evasion Techniques
  • 💻 Hardware Attacks
  • 🔐 Intrusion Techniques
  • 📡 IoT Attacks
  • 🦠 Malware Attacks
  • 🌐 Network Attacks
  • 🌍 Web Application Attacks

---

🗂️ Repository Structure

/
├── Code/
│   ├── dataset_generation.py
│   ├── keywords.txt
├── Dataset/
│   ├── Category sets/
│   │   ├── cysecbench-cloud-attacks.csv
│   │   ├── cysecbench-control-system-attacks.csv
│   │   ├── cysecbench-cryptographic-attacks.csv
│   │   ├── cysecbench-evasion-techniques.csv
│   │   ├── cysecbench-hardware-attacks.csv
│   │   ├── cysecbench-intrusion-techniques.csv
│   │   ├── cysecbench-iot-attacks.csv
│   │   ├── cysecbench-malware-attacks.csv
│   │   ├── cysecbench-network-attacks.csv
│   │   ├── cysecbench-web-application-attacks.csv
│   ├── Full dataset/
│   │   ├── cysecbench.csv
│   ├── Sample sets/
│       ├── cysecbench-500.csv
│       ├── cysecbench-2000.csv
│       ├── cysecbench-6000.csv

---

🚀 Getting Started

⚙️ Prerequisites

• 🐍 Python 3.8+
• 📦 Required libraries: openai (only for dataset generation)

---

📊 Results using CySecBench

🎯 Evaluation Metrics

• ✅ Success Rate (SR): Percentage of prompts bypassing ethical guidelines.
• 📈 Average Rating (AR): Degree of harmfulness in LLM responses (on a scale of 1-5, where 5 is the most harmful).

⚡ Jailbreaking Performance

LLM	Success Rate (SR)	Average Rating (AR)
🤖 Claude	17.4%	2.00
🤖 ChatGPT	65.4%	4.06
🤖 Gemini	88.4%	4.77

---

📜 Citation

If you use CySecBench, please cite:

@article{CySecBench2024,
	title        = {{CySecBench: Generative AI-based CyberSecurity-focused Prompt Dataset for Benchmarking Large Language Models}},
	author       = {Johan Wahréus and Ahmed Mohamed Hussain and Panos Papadimitratos},
	year         = {2025},
	journal      = {arXiv preprint arXiv:2501.01335},
	url          = {https://arxiv.org/abs/2501.01335}
}

⭐ Star This Repository!

If you found CySecBench helpful or interesting, please give this repository a star ⭐ to show your support!

---

🔒 License

This project is licensed under the MIT License. See the LICENSE file for details.