Bypassing the Safety Training of Open-Source LLMs with Priming Attacks

December 19, 2023 · Entered Twilight · 🏛 Tiny Papers @ ICLR

Repo contents: LICENSE, LLAMA_2_LICENSE, README.md, attack_llama.py, attack_vicuna.py, data, environment.yml, few_shot_priming.py, generation_attack.py, get_accuracy.py, install.sh, llama_guard.py, llama_guard, notice.txt

Authors Jason Vega, Isha Chaudhary, Changming Xu, Gagandeep Singh arXiv ID 2312.12321 Category cs.CR: Cryptography & Security Cross-listed cs.AI, cs.CL, cs.LG Citations 42 Venue Tiny Papers @ ICLR Repository https://github.com/uiuc-focal-lab/llm-priming-attacks ⭐ 17 Last Checked 1 month ago

Abstract

With the recent surge in popularity of LLMs has come an ever-increasing need for LLM safety training. In this paper, we investigate the fragility of SOTA open-source LLMs under simple, optimization-free attacks we refer to as $\textit{priming attacks}$, which are easy to execute and effectively bypass alignment from safety training. Our proposed attack improves the Attack Success Rate on Harmful Behaviors, as measured by Llama Guard, by up to $3.3\times$ compared to baselines. Source code and data are available at https://github.com/uiuc-focal-lab/llm-priming-attacks.