In general, CAPTCHA is an image containing distorted letters or numbers. This test involves users typing the results of guessing letters or numbers in the distorted image as security before users can enter or access a website they want to go to. In this paper, a simulation of an automatic reader system for CAPTCHA has been created where the letters and numbers in the image are carried out in several stages. The initial stage involves training using the EMNIST dataset to train the model to recognize general letter and digit characters before recognition in the CAPTCHA image. Furthermore, the process of recognizing letters and numbers in different CAPTCHA images is carried out to read the text contained in the CAPTCHA. The Convolutional Neural Network (CNN) model of the LeNet-5 method is used as a method for reading distorted letters and numbers in CAPTCHA with a high level of accuracy, achieving 88.56%.