AudioJailbreak

News

• [2026.01] AudioJailbreak accepted by IEEE Transactions on Dependable and Secure Computing (TDSC)

Environment setup

git clone https://github.com/ictnlp/LLaMA-Omni
cd LLaMA-Omni

conda create -n llama-omni python=3.10
conda activate llama-omni
pip install pip==24.0
pip install -e .

git clone https://github.com/pytorch/fairseq
cd fairseq
pip install -e . --no-build-isolation

pip install flash-attn --no-build-isolation

Model & dataset preparation

1. Download the Llama-3.1-8B-Omni model from 🤗Huggingface.

2. Download the Whisper-large-v3 model.

import whisper
model = whisper.load_model("large-v3", download_root="models/speech_encoder/")

3. Download the HarmBench-Llama-2-13b-cls from 🤗Huggingface.

4. Download the dataset

• 4.1 Download the dataset AdvBench-harmful_behaviors_custom

• 4.2 unzip

unzip AdvBench-harmful_behaviors_custom.zip

Generate and test jailbreak audio

python generate.py

python get_rsp.py; python eval.py; python analyse.py

We are continuing in improving the code. Stay tuned!

If you find our work and code useful, considering citing it as follows:

@article{audiojailbreak,
  author       = {Guangke Chen and
                  Fu Song and
                  Zhe Zhao and
                  Xiaojun Jia and
                  Yang Liu and
                  Yanchen Qiao and
                  Weizhe Zhang and Weiping Tu and Yuhong Yang and Bo Du},
  title        = {AudioJailbreak: Jailbreak Attacks against End-to-End Large Audio-Language Models},
journal      = {{IEEE} Transactions on Dependable and Secure Computing},
year         = {2026},
}