Why can permutation importance underrate correlated features?

A model relies on either of two strongly correlated variables interchangeably. Why might ordinary one-feature-at-a-time permutation importance assign both variables low importance?