Deep learning models in medical imaging often latch onto shortcut features—spurious correlations such as scanner type, image artifacts, laterality markers, or patient positioning—that allow them to perform well on validation datasets while failing in real clinical settings. This talk explores the mechanisms behind shortcut formation and the risks they pose for safety-critical applications like diagnosis and triage. We review empirical evidence of shortcut reliance across modalities (CT, MRI, X-ray, ultrasound) and examine why standard evaluation pipelines fail to detect them. The talk then surveys current mitigation strategies, including data curation and augmentation, adversarial training, representation disentanglement, and supervised learning. We conclude with practical recommendations for building robust models and designing evaluations that reflect real-world heterogeneity.