World-action models have a fundamental safety gap: they can appear to plan correctly (good future predictions) while actually executing harmful actions, making them vulnerable to adversarial attacks that decouple imagination from execution.
This paper reveals a critical vulnerability in world-action models (WAMs)—AI systems that predict future states while generating robot actions. The authors show that small visual perturbations can cause WAMs to imagine correct futures but execute wrong actions, breaking the safety assumption that internal predictions validate external behavior.