An adversarial attack that causes harmful action shifts while keeping the model's predicted future state realistic and plausible.