Reproduce “FuzzLLM: A Novel and Universal Fuzzing Framework for Proactively Discovering Jailbreak Vulnerabilities in Large Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Dongyu Yao, Jianshu Zhang, Ian G. Harris, Marcel Carlsson
FuzzLLM introduces an automated fuzzing framework for proactively identifying jailbreak vulnerabilities in Large Language Models, demonstrating its ability to uncover flaws in both open-source and commercial models like GPT-3.5-turbo and GPT-4, and generating valuable datasets for safety fine-tuning.
Key insights
- Combo attack classes consistently show greater effectiveness in discovering jailbreak vulnerabilities across various LLMs compared to individual base classes.
- Commercial LLMs like GPT-3.5-turbo and GPT-4, despite advanced defenses, remain vulnerable to fuzzed attacks, albeit generally less so than open-source models.
- FuzzLLM’s automated labeling process, leveraging an LLM, reliably identifies successful jailbreaks, reducing manual effort and providing high-quality data for safety training.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.