This talk will give an overview on the security implications of machine learning approaches and their potential vulnerability to adversarial examples, which could lead to the misclassification of medical events. These insights stem from the BMFTR-funded project SecureNeuroAI. The project aims to develop legally compliant, secure, AI-based solutions for real-time seizure detection from multimodal data, with a special focus on protecting against cyberattacks that could manipulate data or AI workflows and endanger patient safety.