Between Safety Boundaries: Jailbreaking Text-to-Video Models via Temporal Consistency

The researchers propose a jailbreak framework named BSB that exploits the temporal consistency inherent in video generation to attack text-to-video models, achieving an average relative improvement of 18.6% in attack success rate over the strongest baseline on the commercial models evaluated. The method encodes a harmful intent as a transition between two individually harmless boundary states, attacking boundary state pairs that tend to generate unsafe intermediate frames after interpolation. To avoid the high computational cost of direct evaluation in video space, BSB runs Monte Carlo tree search (MCTS) in a cheaper text proxy space and periodically calibrates the search results with sparse video-level evaluations. Experiments are conducted on mainstream commercial text-to-video models including Veo 3.1, Sora 2, Seedance, and Kling v1. The authors point out that temporal consistency is an understudied yet critical attack surface of text-to-video models, and that structured search helps effectively discover vulnerabilities under a limited query budget.