From Compliance to Exploitation: Jailbreak Prompt Attacks on Multimodal LLMs

Large Language Models (LLMs) have seen widespread applications across various domains due to their growing ability to process diverse types of input data, including text, audio, image and video. While LLMs have demonstrated outstanding performance in understanding and generating contexts for different scenarios, they are vulnerable to prompt-based attacks, which are mostly via text input. In this paper, we introduce the first voice-based jailbreak attack against multimodal LLMs, termed as Flanking Attack, which can process different types of input simultaneously towards the multimodal LLMs. Our work is motivated by recent advancements in monolingual voice-driven large language models, which have introduced new attack surfaces beyond traditional text-based vulnerabilities for LLMs. To investigate these risks, we examine the frontier multimodal LLMs, which can be accessed via different types of inputs such as audio input, focusing on how adversarial prompts can bypass its defense mechanisms. We propose a novel strategy, in which the disallowed prompt is flanked by benign, narrative-driven prompts. It is integrated in the Flanking Attack which attempts to humanizes the interaction context and execute the attack through a fictional setting. To better evaluate the attack performance, we present a semi-automated self-assessment framework for policy violation detection. We demonstrate that Flank Attack is capable of manipulating state-of-the-art LLMs into generating misaligned and forbidden outputs, which achieves an average attack success rate ranging from 0.67 to 0.93 across seven forbidden scenarios. These findings highlight both the potency of prompt-based obfuscation in voice-enabled contexts and the limitations of current LLMs' moderation safeguards and the urgent need for advanced defense strategies to address the challenges posed by evolving, context-rich attacks.

SMOClust: Synthetic Minority Oversampling based on Stream Clustering for Evolving Data Streams

Many real-world data stream applications not only suffer from concept drift but also class imbalance. Yet, very few existing studies investigated this joint challenge. Data difficulty factors, which have been shown to be key challenges in class imbalanced data streams, are not taken into account by existing approaches when learning class imbalanced data streams. In this work, we propose a drift adaptable oversampling strategy to synthesise minority class examples based on stream clustering. The motivation is that stream clustering methods continuously update themselves to reflect the characteristics of the current underlying concept, including data difficulty factors. This nature can potentially be used to compress past information without caching data in the memory explicitly. Based on the compressed information, synthetic examples can be created within the region that recently generated new minority class examples. Experiments with artificial and real-world data streams show that the proposed approach can handle concept drift involving different minority class decomposition better than existing approaches, especially when the data stream is severely class imbalanced and presenting high proportions of safe and borderline minority class examples.