Neural networks, despite their robustness in performing complex tasks, exhibit vulnerabilities to adversarial attacks, notably through methods like the Fast Gradient Sign Method (FGSM). FGSM exploits these weaknesses by subtly altering input data to mislead neural networks into making incorrect predictions. The method involves calculating the loss after forward propagation, determining the gradient relative to the image pixels, and then adjusting these pixels to maximize the loss, ultimately tricking the network. Depending on the attacker's knowledge, attacks can be categorized into white box and black box types, with the former providing full access to the model's architecture. The degree of noise added to inputs, controlled by a parameter called epsilon, affects the visibility of these changes and the likelihood of incorrect predictions. By manipulating the gradient directions, FGSM demonstrates how neural networks can be deceived without altering model parameters, highlighting a significant intersection of AI with security challenges.