TEAR: A Time-Aware Automated Red-Teaming Framework for Text-to-Video Models

The researchers propose TEAR, an automated red-teaming framework targeting the temporal dynamic safety risks of text-to-video (T2V) models, to address the shortcoming that existing evaluation methods focus on static images and text generation. The framework trains a time-aware test generator through two-stage optimization (initial generator training and time-aware online preference learning), constructs prompts that are harmless in text, uses temporal dynamics to induce the model to generate policy-violating videos, and uses a refinement model to iteratively improve the prompts' stealthiness and adversarial effectiveness. Experiments show that TEAR achieves an attack success rate exceeding 80% on open-source and commercial T2V systems, significantly higher than the previous best of 57%.