I don't fully understand this. Say we have two features, $$X$$ and $$Y$$, where $$Y=cX + \epsilon$$, that is, Y is correlated with X plus some small amount of random noise. If I run a PCA on this, I will get one large principal component and one small. If the loading vectors correspond to the features themselves, then I will simply get the projection of the original axes onto the principal components; these projected vectors don't seem to have a low angle between them.