Interview question

Handling Imbalanced Classification Datasets

When training a model to predict rare events (e.g., credit card fraud occurring in < 0.1% of transactions), why is standard 'Accuracy' a misleading performance metric, and what metric is generally preferred instead?