Machine learning algorithms are revolutionizing modern society in many domains such as image processing, natural speech understanding, medicine and automotive. In the automotive industry, researchers and developers are actively using machine learning-based approaches for developing automated driving functions. However, before a machine learning algorithm executing safety-related tasks finds its way into series production cars, it has to undergo strict assessment concerning safety. In traditional rule-based programmed software systems established safety engineering processes and practices are successfully applied, whereas data-driven based machine learning algorithms raise new and sometimes obscure safety challenges. This paper describes some of the safety challenges of machine learning in the automotive domain, surveys the topic and suggests potential approaches for addressing the challenges.