Hangtao/badrobot-malicious-queries

BadRobot Malicious Queries

This dataset contains the malicious-query benchmark released with BadRobot: Jailbreaking Embodied LLM Agents in the Physical World. It is intended for research on embodied AI safety, red-teaming, refusal behavior, and safety evaluation of language-model-powered robotic or embodied agents.

The benchmark consists of natural-language requests covering categories such as physical harm, privacy violations, pornography, fraud, illegal activities, hateful conduct, and sabotage. These requests are provided for controlled safety evaluation only.

Dataset Structure

The dataset is provided as a CSV file:

malicious_queries.csv

Columns:

• id: Integer identifier from the original benchmark.
• category: Safety category assigned in the original benchmark.
• request: The malicious request text.
• benchmark: Constant value, badrobot_malicious_queries.
• source_file: Original source filename in the GitHub repository.

Usage

from datasets import load_dataset

dataset = load_dataset("Hangtao/badrobot-malicious-queries", split="train")
print(dataset[0]["request"])

BadRobot Jailbreak Prompts

We also provide jailbreak prompts generated from the malicious-query benchmark using the three BadRobot attack methods. Each row already contains the complete jailbreak prompt in the prompt field.

conceptual = load_dataset("Hangtao/badrobot-malicious-queries", "conceptual_deception_attack", split="train")
contextual = load_dataset("Hangtao/badrobot-malicious-queries", "contextual_jailbreak_attack", split="train")
safety = load_dataset("Hangtao/badrobot-malicious-queries", "safety_misalignment_attack", split="train")

print(conceptual[0]["prompt"])
print(contextual[0]["prompt"])
print(safety[0]["prompt"])

• conceptual_deception_prompts.csv: conceptual-deception prompts.
• contextual_jailbreak_prompts.csv: contextual-jailbreak prompts.
• safety_misalignment_prompts.csv: safety-misalignment prompts.

Source

GitHub repository: https://github.com/Rookie143/BadRobot

Paper: https://arxiv.org/abs/2407.20242

Original file: Physical_Word_Malicious_Queries.xlsx

Supplementary Jailbreak Prompts

The repository also includes jailbreak_prompts.csv, a supplementary collection of 100 recent in-the-wild jailbreak prompts grouped by type. These prompts target LLMs themselves rather than embodied-agent attack strategies, and are provided only for completeness.

Intended Use

This dataset is intended for controlled research on embodied-agent safety, red-teaming, refusal behavior, and safety evaluation.

Out-of-Scope Use

This dataset must not be used to assist real-world harm, automate harmful actions, train systems to execute malicious requests, or bypass safety protections in deployed robotic or embodied systems.

Ethical Considerations

The dataset contains potentially harmful language and requests. Researchers should use it only in controlled environments and should avoid exposing raw examples to end users except when necessary for safety evaluation. Any model outputs generated from these prompts should be reviewed carefully before release.

Citation

@inproceedings{zhangbadrobot,
  title={BadRobot: Jailbreaking Embodied LLM Agents in the Physical World},
  author={Zhang, Hangtao and Zhu, Chenyu and Wang, Xianlong and Zhou, Ziqi and Yin, Changgan and Li, Minghui and Xue, Lulu and Wang, Yichen and Hu, Shengshan and Liu, Aishan and others},
  booktitle={International Conference on Learning Representations (ICLR)},
  year={2025}
}