For the same reason you try to avoid machine learning models from overfitting. If you train a cats vs. dogs machine learning model but all your cat pictures are indoors and all dog pictures are outdoors, the model may inadvertently learn to recognise indoors vs. outdoors instead of learning cat vs. dog features, and will fail miserably on an image of a cat outdoors.

In your case, first of all, it's not 100%. Suppose 90% of attacks are by Xians and 10% are Yians.

Profiling Xians over Yians might mean that you catch 90% of attacks. Maybe 95% if you search some Yians.

On the other hand, learning to properly identify threats, independent of Xian vs. Yian, might allow you to catch 99.9% of attacks.