[guided]We have already used Faltings's theorem to obtain a rational preimage of $u$. The only possible obstruction to surjectivity of the integral map is denominators: the rational preimage might a priori involve a negative power of $l$. We now rule this out.
Extend $u$ to a map on rational Tate modules by defining
\begin{align*}
u_{\mathbb{Q}}: V_lA &\to V_lB, \\
x\otimes c &\mapsto u(x)\otimes c.
\end{align*}
This map is $\mathbb{Q}_l$-linear and $G_K$-equivariant because $u$ is $\mathbb{Z}_l$-linear and $G_K$-equivariant. Faltings's rational theorem gives a unique
\begin{align*}
\alpha \in \operatorname{Hom}_K(A,B)\otimes_{\mathbb{Z}}\mathbb{Q}_l
\end{align*}
with $\Phi_{l,\mathbb{Q}}(\alpha)=u_{\mathbb{Q}}$.
Choose the smallest integer $m \geq 0$ such that $l^m\alpha$ lies in the integral lattice $\operatorname{Hom}_K(A,B)\otimes_{\mathbb{Z}}\mathbb{Z}_l$, and define $\beta := l^m\alpha$. If $m>0$, then the induced Tate-module map satisfies
\begin{align*}
\Phi_l(\beta)=l^m u.
\end{align*}
Since $u(T_lA)\subseteq T_lB$, the image of $l^m u$ is contained in $l^mT_lB$, and because $m>0$ this is contained in $lT_lB$. Therefore the reduction modulo $l$ of $\Phi_l(\beta)$ is the zero map
\begin{align*}
\overline{\Phi_l(\beta)}: T_lA/lT_lA &\to T_lB/lT_lB.
\end{align*}
Now $T_lA/lT_lA$ identifies with $A[l](\overline{K})$, and the same is true for $B$. Under these identifications, the reduction of $\Phi_l(\beta)$ modulo $l$ is exactly the map induced by $\beta$ on $l$-torsion. Thus $\beta$ kills $A[l](\overline{K})$.
We use the external input [Tate module saturation for abelian-variety Hom lattices](/page/Tate%20Module%20Saturation). The theorem says that, for abelian varieties over a field of characteristic $0$, the image of
\begin{align*}
\operatorname{Hom}_K(A,B)\otimes_{\mathbb{Z}}\mathbb{Z}_l
\end{align*}
in $\operatorname{Hom}_{\mathbb{Z}_l}(T_lA,T_lB)$ is $l$-saturated. In the present form, this means: if an element of $\operatorname{Hom}_K(A,B)\otimes_{\mathbb{Z}}\mathbb{Z}_l$ acts as zero on $A[l](\overline{K})$, then it is divisible by $l$ in that $\mathbb{Z}_l$-module. The hypotheses apply because $K$ is a number field, so $\operatorname{char} K=0$, and $A$ and $B$ are abelian varieties over $K$. Applying the saturation theorem to $\beta$ gives
\begin{align*}
\beta = l\gamma
\end{align*}
for some $\gamma \in \operatorname{Hom}_K(A,B)\otimes_{\mathbb{Z}}\mathbb{Z}_l$. But then
\begin{align*}
l^{m-1}\alpha = \gamma,
\end{align*}
which contradicts the minimality of $m$. Hence $m=0$, so the rational preimage $\alpha$ was already integral. Therefore every $u \in \operatorname{Hom}_{G_K}(T_lA,T_lB)$ is in the image of $\Phi_l$.[/guided]