T2VShield: Model-Agnostic Jailbreak Defense for Text-to-Video Models

2025-04-22

AI Summary

Text-to-video (T2V) models are vulnerable to jailbreaking attacks, compromising the safety and reliability of simulation-based applications. To address this, we propose the first generic defense framework for black-box T2V models—requiring no access to model parameters or internal architecture. Our method comprises three core components: (1) a novel prompt purification mechanism integrating semantic reasoning and cross-modal retrieval to rewrite malicious inputs; (2) a temporal–modal joint multi-scale consistency detection module that identifies anomalies across spatiotemporal, text–video, and audio–video dimensions; and (3) the first human-centered audiovisual safety evaluation protocol. Evaluated on five mainstream T2V platforms, our framework reduces jailbreaking success rates by up to 35%. It supports both open- and closed-source models in a plug-and-play manner, significantly enhancing perceptual-level safety and trustworthiness of generated content.

Technology Category

• Computer Vision: Large Vision Models
• Natural Language Processing: Safety and Robustness
• Machine Learning: Large Multimodal Models (LMMs)

Application Category

• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Security and Privacy: Large-scale security measurements
• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG

Abstract

The rapid development of generative artificial intelligence has made text to video models essential for building future multimodal world simulators. However, these models remain vulnerable to jailbreak attacks, where specially crafted prompts bypass safety mechanisms and lead to the generation of harmful or unsafe content. Such vulnerabilities undermine the reliability and security of simulation based applications. In this paper, we propose T2VShield, a comprehensive and model agnostic defense framework designed to protect text to video models from jailbreak threats. Our method systematically analyzes the input, model, and output stages to identify the limitations of existing defenses, including semantic ambiguities in prompts, difficulties in detecting malicious content in dynamic video outputs, and inflexible model centric mitigation strategies. T2VShield introduces a prompt rewriting mechanism based on reasoning and multimodal retrieval to sanitize malicious inputs, along with a multi scope detection module that captures local and global inconsistencies across time and modalities. The framework does not require access to internal model parameters and works with both open and closed source systems. Extensive experiments on five platforms show that T2VShield can reduce jailbreak success rates by up to 35 percent compared to strong baselines. We further develop a human centered audiovisual evaluation protocol to assess perceptual safety, emphasizing the importance of visual level defense in enhancing the trustworthiness of next generation multimodal simulators.

Problem

• Defends text-to-video models against jailbreak attacks
• Addresses semantic ambiguities in malicious prompts
• Detects harmful content in dynamic video outputs

Innovation

• Prompt rewriting using reasoning and multimodal retrieval
• Multi-scope detection for local and global inconsistencies
• Model-agnostic defense without internal parameter access