Interesting taxonomy of machine-learning failures (pdf) that encompasses both mistakes and attacks, or — in their words — intentional and unintentional failure modes. It’s a good basis for threat modeling.