Autoencoders (AE) have recently been widely employed to approach the novelty\ndetection problem. Trained only on the normal data, the AE is expected to\nreconstruct the normal data effectively while fail to regenerate the anomalous\ndata, which could be utilized for novelty detection. However, in this paper, it\nis demonstrated that this does not always hold. AE often generalizes so\nperfectly that it can also reconstruct the anomalous data well. To address this\nproblem, we propose a novel AE that can learn more semantically meaningful\nfeatures. Specifically, we exploit the fact that adversarial robustness\npromotes learning of meaningful features. Therefore, we force the AE to learn\nsuch features by penalizing networks with a bottleneck layer that is unstable\nagainst adversarial perturbations. We show that despite using a much simpler\narchitecture in comparison to the prior methods, the proposed AE outperforms or\nis competitive to state-of-the-art on three benchmark datasets.\n