The Price of Consistency: Exploiting Visual Anchors for Multimodal Jailbreaking in Video Generation

Modern video generation has shifted from purely text-driven models to multi-conditional architectures that utilize reference images to ensure spatiotemporal consistency. While these reference images significantly improve the quality and controllability of generated videos, their impact on safety has remained largely unexamined. This paper investigates how these visual anchors reshape the safety behavior of video generation models, revealing a fundamental trade-off between consistency and security.

The researchers identify the visual anchoring effect, a mechanism where the reference image acts as a persistent constraint throughout the generation process. In text-only models, safety alignment often relies on semantic drift—the model's tendency to steer away from harmful prompts toward benign content. However, the visual anchor forces the model to remain faithful to the initial frame, effectively suppressing this drift and trapping the model within a narrow generative space that includes the harmful intent. The authors term this phenomenon the price of consistency.

To exploit this vulnerability, the authors propose Decoupling Intent via Visual Anchors (DIVA), a training-free jailbreak framework. DIVA decomposes a harmful intent into a static visual anchor image and a dynamic motion text prompt. By employing a dual-criteria selection mechanism, the framework balances the stealthiness of the input (to bypass safety filters) with the semantic preservation of the original harmful intent. This approach allows for the systematic generation of adversarial image-text pairs that force the victim model to reconstruct harmful content.

Extensive experiments across five open-source models and three leading commercial platforms demonstrate that DIVA achieves a significantly higher Attack Success Rate (ASR) than existing text-only jailbreak methods. The authors also introduce TI2VSafetyBench, the first comprehensive safety benchmark for multi-conditional video generation, to facilitate future research into these overlooked vulnerabilities. The study highlights the urgent need for more robust safety alignment strategies that account for the unique risks introduced by multi-conditional generation paradigms.