Jailbreak attacks cause large language models (LLMs) to generate harmful, unethical, or otherwise
unwanted content. Evaluating these attacks presents a number of challenges, and the current
landscape of benchmarks and evaluation techniques is fragmented. First, assessing whether LLM
responses are indeed harmful requires open-ended evaluations which are not yet standardized.
Second, existing works compute attacker costs and success rates in incomparable ways. Third,
some works lack reproducibility as they withhold adversarial prompts or code, and rely on changing
proprietary APIs for evaluation. Consequently, navigating the current literature and tracking
progress can be challenging.
To address this, we introduce JailbreakBench, a centralized benchmark with the following components:
- Repository of jailbreak artifacts. An evolving dataset of state-of-the-art
adversarial prompts at https://github.com/JailbreakBench/artifacts, referred to as jailbreak artifacts, which are explicitly required for submissions
to our benchmark to ensure reproducibility.
- Standardized evaluation framework. Our library at https://github.com/JailbreakBench/jailbreakbench that includes a clearly defined threat model, system prompts, chat templates, and scoring functions.
- Leaderboard. Our leaderboards here (https://jailbreakbench.github.io/) that track the performance of attacks and defenses for various LLMs.
- Dataset. A representative dataset named JBB-Behaviors at https://huggingface.co/datasets/JailbreakBench/JBB-Behaviors composed of 100 distinct misuse behaviors (with 55% original examples and the rest sourced from AdvBench and TDC/HarmBench) divided into ten broad categories corresponding to OpenAI's usage policies. Moreover, now it is complemented with 100 benign behaviors that can be used to quickly evaluate overrefusal rates for new models and defenses.
We have carefully considered the potential ethical implications of releasing this benchmark,
and believe that it will be a net positive for the community. Our jailbreak artifacts can expedite
safety training for future models. Over time, we will expand and adapt the benchmark to reflect
technical and methodological advances in the research community.
Leaderboard: Open-Source Models
Leaderboard: Closed-Source Models
Contribute to JailbreakBench
We welcome contributions in terms of both new attacks and defenses. Please check
here and
here
for more details. Feel free to contact us for any questions or suggestions.
If you use the JBB-Behaviors dataset, we ask you consider citing the following works:
Moreover, consider citing our whitepaper if you use the datasets, reference our leaderboard, or if you are using our evaluation library:
@inproceedings{chao2024jailbreakbench,
title={JailbreakBench: An Open Robustness Benchmark for Jailbreaking Large Language Models},
author={Patrick Chao and Edoardo Debenedetti and Alexander Robey and Maksym Andriushchenko and Francesco Croce and Vikash Sehwag and Edgar Dobriban and Nicolas Flammarion and George J. Pappas and Florian Tramèr and Hamed Hassani and Eric Wong},
booktitle={NeurIPS Datasets and Benchmarks Track},
year={2024}
}