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

TL;DR

This paper demonstrates that simple priming attacks can effectively bypass safety training in open-source large language models, significantly increasing the success rate of harmful behavior elicitation.

Contribution

It introduces priming attacks as a new, simple method to bypass safety alignment in open-source LLMs, showing their effectiveness without optimization.

Findings

Priming attacks increase attack success rate by up to 3.3 times.

Simple, optimization-free attacks can bypass safety training.

Source code and data are publicly available.

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.