Abstract
This paper examines the dual role of generative artificial intelligence in cybersecurity: as an enabler of attacks and as a practical defense component. It is shown that realistic generation of text, images, and audio lowers the barrier for abuse (phishing, deepfakes, impersonation), while machine learning systems face a distinct risk classsmall but targeted input perturbations. The practical part is implemented as a reproducible benchmark around an MNIST handwritten-digit classifier. After verifying stable performance on clean inputs, adversarial examples were generated using FGSM with multiple values of e. The results confirm that a visually subtle perturbation can flip a confident prediction to an incorrect class (in a representative case, a “7” was forced to be classified as “3”). Two reconstruction-based defenses were then compared: a standard autoencoder (AE) and a denoising autoencoder (DAE) trained to recover clean signals from noisy inputs. As e increases, DAE reduces the misclassification rate more noticeably than AE, which is consistent with its training objective of separating high-frequency noise from meaningful strokes. To make the defense actionable in an operational setting, a simple anomaly detector based on reconstruction energy E(x) was also introduced. This adds a second layer: the DAE attempts to restore the input for correct classification, while E(x) provides an explicit alert signal suitable for logging and incident review.