StyleBreak: Revealing Alignment Vulnerabilities in Large Audio-Language Models via Style-Aware Audio Jailbreak

TL;DR

This paper introduces StyleBreak, a style-aware audio jailbreak framework that uncovers vulnerabilities in large audio-language models by systematically manipulating speech attributes, revealing critical security weaknesses.

Contribution

The paper presents a novel style-aware attack method and a query-adaptive policy network to improve the effectiveness and efficiency of audio jailbreaks on LAMs.

Findings

LAMs are vulnerable to diverse speech attribute manipulations

StyleBreak significantly improves attack success rates

The framework highlights urgent need for robust LAM alignment

Abstract

Large Audio-language Models (LAMs) have recently enabled powerful speech-based interactions by coupling audio encoders with Large Language Models (LLMs). However, the security of LAMs under adversarial attacks remains underexplored, especially through audio jailbreaks that craft malicious audio prompts to bypass alignment. Existing efforts primarily rely on converting text-based attacks into speech or applying shallow signal-level perturbations, overlooking the impact of human speech's expressive variations on LAM alignment robustness. To address this gap, we propose StyleBreak, a novel style-aware audio jailbreak framework that systematically investigates how diverse human speech attributes affect LAM alignment robustness. Specifically, StyleBreak employs a two-stage style-aware transformation pipeline that perturbs both textual content and audio to control linguistic, paralinguistic, and extralinguistic attributes. Furthermore, we develop a query-adaptive policy network that automatically searches for adversarial styles to enhance the efficiency of LAM jailbreak exploration. Extensive evaluations demonstrate that LAMs exhibit critical vulnerabilities when exposed to diverse human speech attributes. Moreover, StyleBreak achieves substantial improvements in attack effectiveness and efficiency across multiple attack paradigms, highlighting the urgent need for more robust alignment in LAMs.