Weak-to-Strong Jailbreaking on Large Language Models

📣 Update: Our paper has been accepted to ICML 2025!

📄 arXiv | 🤗 HuggingFace Paper Page

Overview

Despite major advances in aligning large language models (LLMs), red-teaming efforts consistently reveal vulnerabilities: even well-aligned LLMs can be jailbroken to produce harmful outputs via adversarial prompts, fine-tuning, or decoding tricks.

This repository implements Weak-to-Strong Jailbreaking — a novel and efficient inference-time attack that leverages small (7B) unsafe/aligned LLMs to guide the generation of much larger (e.g., 70B) aligned models into producing unsafe outputs. Surprisingly, the attack only requires one forward pass through each small model, making it both computationally cheap and highly effective.

Key Insight

Aligned and jailbroken LLMs mainly diverge in their initial decoding steps. This enables us to apply log-probability algebra — using small models to shift the strong model's token distribution early in generation — resulting in high attack success rates (ASR > 99%) with minimal cost.

Pipeline Illustration

We summarize the trade-offs of different jailbreaking strategies below:

Repository Structure

• data/: Contains the data used for the experiments.
• run.py: Contains the scripts used to run the experiments.
• generate.py: Contains the scripts used to generate the results.
• eval_asr.py: Contains the scripts used to evaluate the attack success rate.
• eval_gpt.py: Contains the scripts used to evaluate the GPT4 scores.
• eval_harm.py: Contains the scripts used to evaluate the Harm scores.

For getting the unsafe small model, please refer to this repo: https://github.com/BeyonderXX/ShadowAlignment

Running the experiments

python run.py --beta 1.50 --batch_size 16 --output_file "[OUTPUT FILE NAME]" --att_file "./data/advbench.txt'

Need to confige the bad model path in run.py firstly.

Evaluating the results

Find the examples in eval_asr.py, eval_gpt.py, and eval_harm.py to evaluate the results.

Citation

If you find the code useful, please cite the following paper:

@article{zhao2024weak,
  title={Weak-to-Strong Jailbreaking on Large Language Models},
  author={Zhao, Xuandong and Yang, Xianjun and Pang, Tianyu and Du, Chao and Li, Lei and Wang, Yu-Xiang and Wang, William Yang},
  journal={arXiv preprint arXiv:2401.17256},
  year={2024}
}