If they are building an algorithm which overfits to training data, their algorithm will lose performance on in the actual validation test.

It is NOT like looking at leaked answers. It's like re-reading the textbook more than everyone else. Definitely an unfair advantage however.