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