Images easily fool convolutional neural networks with noise. They are not as secure as previously thought. We show this by training a VGG16 model on the Mapillary data set on over 200,000 images. We were able to trick this model into a classification of images that appear to be random noise as 100% confidence prediction of a particular class. Ultimately, we found these ideal images as a result of a random search over the image space. With our small images 32x32x3, we were able to find these relatively quickly and prove the drawbacks of convolutional neural networks and their security.