Cyber security system could benefit from the deep learning breakthrough. However, there are, today, critical issues with deep learning. The major issue is the lack of robustness of classical deep networks. This lack of robustness allows to compute well documented adversarial examples: confusing a network with just the addition of a low amplitude perturbation. Here, I show that this lack of robustness also allows adversarial model: possibility to change the global model behaviour with just the addition of a low amplitude perturbation to the train set. This property could be a real security faults for continuously trained system. Evaluation of the efficiency of adversarial model is provided on image classification datasets.