In 2013 a machine learning competition offered a prize for the most accurate detection of right whale calls based on audio data. The organizers soon discovered data leakage problems in the first release of the dataset, and this article explains what happened. It's a short but interesting article that serves as an excellent example of how subtle or not-so-subtle leakage can occur in specific features.

https://www.kaggle.com/c/the-icml-2013-whale-challenge-right-whale-redux/discussion/4865#25839#post25839