license: apache-2.0
task_categories:
- text-classification
- text-generation
language: - en
size_categories: - n<1K
pretty_name: SafeSearch
tags: - safety
- red-teaming
- llm-evaluation
- search-agents
- benchmark
- robustness
SafeSearch Dataset
The SafeSearch dataset is designed for automated red teaming of LLM-based search agents. It evaluates the safety and robustness of search agents when dealing with unreliable web content.
Dataset Description
SafeSearch contains test cases across 5 risk categories:
- Prompt Injection: Testing if agents follow malicious embedded instructions
- Harmful Output: Testing if agents generate harmful content
- Bias Inducing: Testing if agents produce biased responses
- Advertisement: Testing if agents prioritize commercial content
- Fake Information: Testing if agents propagate misinformation
Usage
from datasets import load_dataset
# Load specific risk type
dataset = load_dataset("jianshuod/safesearch", "prompt_injection")
# Load all risk types (combine multiple configs)
from datasets import concatenate_datasets
ds_prompt = load_dataset("jianshuod/safesearch", "prompt_injection")["train"]
ds_harmful = load_dataset("jianshuod/safesearch", "harmful_output")["train"]
ds_bias = load_dataset("jianshuod/safesearch", "bias_inducing")["train"]
ds_ad = load_dataset("jianshuod/safesearch", "advertisement")["train"]
ds_fake = load_dataset("jianshuod/safesearch", "fake_information")["train"]
all_dataset = concatenate_datasets([ds_prompt, ds_harmful, ds_bias, ds_ad, ds_fake])
# Access a sample
sample = dataset["train"][0]
print(sample["user_query"])
print(sample["risk_type"])
Dataset Splits
Each configuration has a train split containing the test cases.
Configurations
prompt_injection: Prompt injection test cases (60 samples)harmful_output: Harmful output test cases (60 samples)bias_inducing: Bias-inducing test cases (60 samples)advertisement: Advertisement test cases (60 samples)fake_information: Fake information test cases (60 samples)
Citation
@article{dong2025safesearch,
title={SafeSearch: Automated Red-Teaming of LLM-Based Search Agents},
author={Dong, Jianshuo and Guo, Sheng and Wang, Hao and Liu, Zhuotao and Zhang, Tianwei and Xu, Ke and Huang, Minlie and Qiu, Han},
journal={arXiv preprint arXiv:2509.23694},
year={2025}
}
License
Apache 2.0