Human motion prediction has achieved a brilliant performance with the help of CNNs, which facilitates human-machine cooperation. However, currently, there is no work evaluating the potential risk in human motion prediction when facing adversarial attacks, which may cause danger in real applications. The adversarial attack will face two problems against human motion prediction: 1. For naturalness, pose data is highly related to the physical dynamics of human skeletons where Lp norm constraints cannot constrain the adversarial example well; 2. Unlike the pixel value in images, pose data is diverse at scale because of the different acquisition equipment and the data processing, which makes it hard to set fixed parameters to perform attacks. To solve the problems above, we propose a new adversarial attack method that perturbs the input human motion sequence by maximizing the prediction error with physical constraints. Specifically, we introduce a novel adaptable scheme that facilitates the attack to suit the scale of the target pose and two physical constraints to enhance the imperceptibility of the adversarial example. The evaluating experiments on three datasets show that the prediction errors of all target models are enlarged significantly, which means current convolution-based human motion prediction models can be easily disturbed under the proposed attack. The quantitative analysis shows that prior knowledge and semantic information modeling can be the key to the adversarial robustness of human motion predictors. The qualitative results indicate that the adversarial sample is hard to be noticed when compared frame by frame but is relatively easy to be detected when the sample is animated.