Data Leakage
The model having access at training time to information it will not have when it makes a real prediction.
One definition covers every variety, and the named types are just routes across that line: a feature computed from the outcome, a window reaching past the prediction moment, the same entity on both sides of a split, a fitted transform fitted before splitting, or a duplicate row in train and test.
What makes it the nastiest bug class in applied work is the direction of the symptom. A crash announces itself and a weak feature shows up as a flat score, but leakage raises your validation number, so nothing in your metrics can detect it.
Diagnose it by asking of every column whether its value was knowable at the prediction instant, then re-splitting by time and by group and seeing whether the score collapses. The most common single instance is a scaler, imputer or encoder fitted before the split rather than inside a pipeline.