Researchers have developed a new framework called BSB to exploit temporal consistency in text-to-video (T2V) models for jailbreaking. This method encodes harmful intent as transitions between safe boundary states, which are then interpolated to generate unsafe intermediate frames. BSB utilizes Monte Carlo Tree Search in a textual proxy space, calibrated with sparse video evaluations, to efficiently discover vulnerabilities in black-box scenarios. Experiments on models like Veo 3.1 and Sora 2 demonstrated BSB's superior performance over existing jailbreak methods, achieving an average 18.6% relative gain in attack success rate. AI
IMPACT This research highlights a new attack vector for T2V models, potentially impacting safety measures and content moderation efforts.
RANK_REASON Academic paper detailing a new method for jailbreaking text-to-video models. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
