Common instances: restated financials used on the original announcement date; quarterly data applied before its reporting lag; index membership applied before it was announced.
The subtle one. Standardising a feature using the mean and standard deviation of the entire dataset leaks future information into every observation. This is extremely common and rarely noticed, because nothing about the code looks wrong.
The tell. A backtest with suspiciously smooth performance, or a Sharpe that collapses when you re-run with proper point-in-time data.