This article explores the challenges and solutions for enhancing machine learning (ML) safety in autonomous vehicles. It examines the gaps in current automotive safety standards when applied to ML systems and proposes practical solutions to improve reliability and safety. The discussion is organized around two key strategies: implementing robust error detection mechanisms for safe failure modes, and improving algorithm robustness to enhance safety margins across various operational conditions. The article presents concrete implementations of these strategies, including a student model for predicting failures in steering control, an out-of-distribution sample detector, and a cross-domain object detection model for UAVs. Additionally, it outlines future research directions in security against adversarial attacks, procedural safeguards for user experience, and the need for interdisciplinary collaboration to address the complex challenges of ML safety in autonomous vehicles.