language:
- en
license: cc-by-4.0
pretty_name: JailbreakDB
task_categories: - text-classification
tags: - llm-safety
- jailbreak
- prompts
- security
configs: - config_name: prompt_corpus
default: true
data_files:- split: jailbreak
path: text_jailbreak_unique.csv - split: regular
path: text_regular_unique.csv
- split: jailbreak
JailbreakDB
JailbreakDB is a large-scale prompt corpus for LLM safety and prompt-security
research. It contains two deduplicated, text-only CSV files:
text_jailbreak_unique.csv(~6.6M rows): jailbreak and adversarial prompts.text_regular_unique.csv(~5.7M rows): benign prompts.
Each row stores the prompt text and lightweight source metadata. The associated
PromptSecurity evaluation measurements are maintained as a separate Hugging Face
dataset: https://huggingface.co/datasets/youbin2014/PromptSecurity-Eval.
Related resources:
- Code: https://github.com/datasec-lab/PromptSecurity
- Leaderboard: https://datasec-lab.github.io/PromptSecurityLeaderboard/
Data Files
| Split | File | Description |
|---|---|---|
jailbreak |
text_jailbreak_unique.csv |
Jailbreak and adversarial prompts. |
regular |
text_regular_unique.csv |
Benign prompts. |
CSV Field Reference
The prompt-corpus CSV files use the following fields:
| Field | Meaning |
|---|---|
system_prompt |
Optional system-side instruction associated with the prompt. |
user_prompt |
User-side prompt text. |
jailbreak |
Binary prompt label; 1 denotes jailbreak/adversarial prompt and 0 denotes benign prompt. |
source |
Source corpus or collection from which the prompt was collected. |
tactic |
Prompt tactic or category when available. |
Loading
from datasets import load_dataset
corpus = load_dataset("youbin2014/JailbreakDB", "prompt_corpus")
print(corpus)
The CSV files can also be loaded directly:
from datasets import load_dataset
files = {
"jailbreak": "hf://datasets/youbin2014/JailbreakDB/text_jailbreak_unique.csv",
"regular": "hf://datasets/youbin2014/JailbreakDB/text_regular_unique.csv",
}
corpus = load_dataset("csv", data_files=files)
print(corpus)
Paper
If you use JailbreakDB, please cite the PromptSecurity paper accepted at the
2027 IEEE Symposium on Security and Privacy (S&P / Oakland):
@inproceedings{hong2027sokpromptsecurity,
title = {{SoK}: Systematizing {LLM} Prompt Security: Taxonomies, Datasets, and Unified Evaluation of Attacks and Defenses},
author = {Hong, Hanbin and Wu, Shuang and Feng, Shuya and Naderloui, Nima and Yan, Shenao and Zhang, Jingyu and Arastehfard, Ali and Huang, Heqing and Hong, Yuan},
booktitle = {2027 IEEE Symposium on Security and Privacy (SP)},
year = {2027},
note = {To appear}
}
Safety Notice
This dataset may contain harmful, offensive, or disturbing prompts. It is
intended strictly for research on model safety, jailbreak robustness, defense
evaluation, and prompt-security analysis. Please review your institutional and
legal requirements before use.