ChatBug: A Common Vulnerability of Aligned LLMs Induced by Chat Templates

TL;DR

This paper uncovers a vulnerability called ChatBug in aligned LLMs caused by chat templates, showing how malicious prompts can bypass safety measures, and discusses mitigation strategies with trade-offs.

Contribution

The paper identifies ChatBug, a novel vulnerability in LLMs induced by chat templates, and develops attacks demonstrating its exploitation, highlighting challenges in balancing safety and performance.

Findings

ChatBug can be exploited to bypass safety in 8 SOTA LLMs

Existing jailbreaks are more effective when combined with ChatBug

Adversarial training mitigates ChatBug but reduces model performance

Abstract

Large language models (LLMs) are expected to follow instructions from users and engage in conversations. Techniques to enhance LLMs' instruction-following capabilities typically fine-tune them using data structured according to a predefined chat template. Although chat templates are shown to be effective in optimizing LLM performance, their impact on safety alignment of LLMs has been less understood, which is crucial for deploying LLMs safely at scale. In this paper, we investigate how chat templates affect safety alignment of LLMs. We identify a common vulnerability, named ChatBug, that is introduced by chat templates. Our key insight to identify ChatBug is that the chat templates provide a rigid format that need to be followed by LLMs, but not by users. Hence, a malicious user may not necessarily follow the chat template when prompting LLMs. Instead, malicious users could leverage their knowledge of the chat template and accordingly craft their prompts to bypass safety alignments of LLMs. We develop two attacks to exploit the ChatBug vulnerability. We demonstrate that a malicious user can exploit the ChatBug vulnerability of eight state-of-the-art (SOTA) LLMs and effectively elicit unintended responses from these models. Moreover, we show that ChatBug can be exploited by existing jailbreak attacks to enhance their attack success rates. We investigate potential countermeasures to ChatBug. Our results show that while adversarial training effectively mitigates the ChatBug vulnerability, the victim model incurs significant performance degradation. These results highlight the trade-off between safety alignment and helpfulness. Developing new methods for instruction tuning to balance this trade-off is an open and critical direction for future research