World models and video understanding models are increasingly used in dynamic visual environments, but their resilience to physical adversarial attacks has received little attention. A new preprint on arXiv investigates this gap by asking whether a single physical perturbation can generalize across trajectories — that is, remain effective as the viewpoint and time shift.

The authors frame this as a trajectory-generalizable attack problem, distinct from single-frame or single-viewpoint attacks. While the abstract does not disclose specific results, the stated goal is to determine whether one physical change can fool models across a range of conditions, which would represent a serious practical threat.

Because the paper is a preprint and the abstract is truncated, the full methodology and findings are not yet available. The significance lies in the question itself: if such attacks succeed, they would expose a new class of vulnerability for models operating in real-world, dynamic scenes.