This study presents a rigorously developed framework for the detection of harmful visual stimuli within the context of cognitive cybersecurity. A convolutional neural network (CNN), based on a fine-tuned ResNet50 backbone, is employed for binary image classification. The dataset was constructed by integrating participant-provided subjective annotations with curated image repositories, followed by the application of the isolation forest algorithm to eliminate outliers and ensure dataset consistency. Model training involved k-fold cross-validation, resulting in consistently high precision and recall across folds, despite inherent variability in individual emotional responses. To interpret the model’s decision-making process, Grad-CAM visualizations were utilized, highlighting the most influential regions for the classification decision. Furthermore, a comparative evaluation against OpenAI’s universal multimodal moderation model was conducted. The results indicate that the proposed domain-specific approach outperforms the general-purpose baseline in the context of the targeted threat scenarios and user population. These findings underscore the importance of context-aware model design and contribute to the advancement of adaptive, user-centric moderation systems. The proposed methodology lays the groundwork for further research in automated content safety and the development of personalized cybersecurity interventions.