**One can't exhibit the counter example asked**. Applying a secure PRF $F$ to the output of a collision-resistant function $H$ yields a secure MAC (only at worst slightly less secure than the weakest of $F$ and $H$).
<br><sub>That holds a fortiori if $H$ is already a secure MAC, as in the title, rather than merely collision-resistant as in the question as disambiguated by [comment][1].</sub>

Proof sketch: distinguishing $F(H(m))$ from random for chosen distinct $m$ requires breaking the distinguishability of $F$ or finding a collision for $H$.


  [1]: https://crypto.stackexchange.com/q/59201/555#comment130151_59201