Machine learning (ML) algorithms allow computers to learn without being explicitly programmed. Their utilization is spreading to highly sophisticated tasks across multiple domains, like medical diagnostics or fully autonomous vehicles. Whereas this presents a great potential, it also raises new safety concerns as ML has many specificities that make its behaviour prediction and assessment much different from explicitly programmed software systems. A good number of relevant questions are open topics of research. This paper surveys the published work in the area, covering both supervised and reinforcement learning. It presents an overview of the characteristics of ML that (safety) engineers should understand, the challenges posed by them in a safety relevant context, and potential approaches for addressing such challenges. To ML practitioners less familiar with safety concerns, the paper offers a wide-ranging introduction to the subject.