As Yves points out, the answer is 'no' if you replace 'malnormal' by 'almost malnormal'.  However, the answer to your first question is 'yes': it is the case that the Karrass--Solitar result remains true if you only assume that $H$ is malnormal in $A$.

The action of a group $G$ on a tree $T$ is called *$k$-acylindrical* if, for every $g\neq 1$, the diameter of the fixed-point set of $g$ is at most $k$.  If $H$ is malnormal in both $A$ and $B$ then the action on the Bass--Serre tree is 1-acylindrical; if it's just malnormal in $A$, the action is 2-acylindrical.

Easy exercise: If $T$ is $k$-acylindrical for some $k$ and $g$ acts hyperbolically, then the centralizer of $g$ is infinite cyclic.

So, after conjugating, we're left with the case in which $g$ is in $A$ or $B$.   Note that the centralizer $Z(g)$ maps $\mathrm{Fix}(g)$ to itself.  If $g\in A$ and is not conjugate into $H$ then $\mathrm{Fix}(g)$ is a single vertex stabilized by $A$, so $Z(g)\smallsetminus A$ as required.  On the other hand, if $g\in B$ then $\mathrm{Fix}(g)$ is contained in a tree of diameter at most 2, with the central vertex stabilized by $B$.  Every automorphism of $\mathrm{Fix}(g)$ with no edge inversions fixes the central vertex, and it follows that $Z(g)$ is indeed contained in $B$, as required.