license: mit
language:
- en
size_categories: - 1K<n<10K
HarDBench(Harmful Draft Benchmark) : A Benchmark for Draft-Based Co-Authoring Jailbreak Attacks for Safe Human–LLM Collaborative Writing
⚠️ WARNING: This dataset contains high-risk content, including detailed instructions and examples for the manufacturing of illegal substances (drugs), the construction of explosives and weapons, and the execution of cyberattacks. This data was curated specifically for "Red Teaming" and AI safety benchmarks to help models refuse dangerous requests. It is intended STRICTLY for research and defensive purposes. Unauthorized use or application of this data poses significant real-world safety and security risks. Users must exercise extreme caution.
Paper: A Benchmark for Draft-Based Co-Authoring Jailbreak Attacks for Safe Human–LLM Collaborative Writing
Project Page: [Project Page]
About
HarDBench (Harmful Draft Benchmark) is a systematic benchmark designed to evaluate the robustness of Large Language Models (LLMs) against draft-based co-authoring jailbreak attacks.
Access & Usage Policy
Although this dataset is released under the MIT License, the sensitive nature of the content requires all users to adhere to the following Ethical Guidelines:
- Research Only: This data must be used exclusively for research, academic, or educational purposes related to AI safety and ethics.
- Strict Non-Redistribution: You may not host, mirror, or redistribute this dataset (or any derivative containing the raw harmful prompts) to any third party without explicit permission.
- No Malicious Use: Do not use this data to train models for generating harmful content or to facilitate illegal activities.
- Safety Protocols: Ensure that any model outputs generated using this data are handled in a secure environment to prevent accidental exposure to end-users.
Dataset Structure
The dataset is organized into multiple configurations to test different safety dimensions:
- prompt: The input query designed to test model boundaries.
- context: (If applicable) The background information or persona provided for the attack.
- category: The type of harm being tested (e.g., Cyberattacks, Harassment, Misinformation).
License
The code and metadata are licensed under the MIT License. However, users are bound by the safety commitment made during the access request (Gating) process.
Citation
If you use this dataset in your research, please cite it as follows:
@inproceedings{kim2026hardbench,
title={HarDBench: A Benchmark for Draft-Based Co-Authoring Jailbreak Attacks for Safe Human--LLM Collaborative Writing},
author={Kim, Euntae and Han, Soomin and Chang, Buru},
booktitle={Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
pages={40785--40831},
year={2026}
}