Title: Exploiting Multi-Clip Video Inputs to Jailbreak Multimodal Large Language Models

Abstract:

The rapid evolution of multimodal large language models (MLLMs) in their ability to process video has raised significant alarms regarding their susceptibility to malicious exploitation. While previous research has demonstrated that safety alignments within MLLMs can be circumvented via visual data, the specific characteristics of video inputs that trigger these vulnerabilities remain poorly understood. To bridge this knowledge gap, we present Multi-Clip Video (MCV) SafetyBench, a comprehensive dataset comprising 2,920 videos. This resource is designed to assess how the variety of video content influences MLLM security. Each entry in the dataset features multiple short clips that illustrate a range of contexts pertinent to a specific harmful query.

Testing across eight prominent video MLLMs revealed a consistent trend: the success rate of attacks rises as the number of clips increases. Our analysis highlights three key vulnerabilities inherent to the video modality: it is more prone to exploitation than the image modality; dynamic videos pose a greater risk than static ones; and models are less secure when videos incorporate a wider array of contexts. Drawing on these insights, we propose a defensive approach that capitalizes on the relative stability of image-based inputs.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC