SpeechJBB: Probing Safety Alignment and Comprehension in Large Audio Language Models under Code-Switched Speech

The article introduces SpeechJBB, a novel audio jailbreak dataset designed to evaluate the safety alignment of large audio language models (LALMs) under code-switched speech conditions. The dataset facilitates benchmarking across various state-of-the-art LALMs, revealing that code-switched harmful audio leads to significantly high jailbreak success rates, particularly with non-English pairs. This research highlights critical vulnerabilities in LALMs' safety mechanisms, emphasizing the need for improved robustness against multilingual and obfuscated prompts in real-world applications.