Consider the CPTP map $\Phi: M_n(\mathbb{C})\rightarrow M_n(\mathbb{C})$ that is not injective, which means the null space of $\Phi$ is non-trivial. My main question is, what can we say about the matrices in this null space?
By this question, we can take a basis of $M_n(\mathbb{C})$, denoted by $\{D_1,D_2,\dots,D_{n^2}\}$ where $D_i$'s are all $n\times n$ density matrices. Since $\Phi$ is not injective, we can get that $\operatorname{span}\{\Phi(D_1),\Phi(D_2),\dots,\Phi(D_{n^2})\}$ is a proper subspace of $M_n(\mathbb{C})$. In other words, $\Phi(D_1),\Phi(D_2),\dots,\Phi(D_{n^2})$ are not linearly independent. So we have that $$ \Phi(\alpha_1 D_1 + \alpha_2 D_2 + \cdots \alpha_{n^2} D_{n^2})=\alpha_1 \Phi(D_1) + \alpha_2 \Phi(D_2) + \cdots + \alpha_{n^2} \Phi(D_{n^2})=0 $$ for some $\alpha_i$'s that are not all zero. On the other hand, since $D_1,D_2,\dots,D_{n^2}$ are linearly independent, this implies that $\alpha_1 D_1 + \alpha_2 D_2 + \cdots \alpha_{n^2} D_{n^2}$ is a non-trivial element in the null space of $\Phi$.
Now let's look at this non-trivial element. It is clearly Hermitian, and its trace is zero. Of course, we can also use it to span a one-dimensional subspace of the null space. Is that all we can say about the matrices in the null space of $\Phi$?
Edit: This proof may be not sufficient. But the above conclusion should be still correct, by the proof that I provided in the comments.