As soon as $n=2$, and $A$ is a symmetric matrix, the answer is surely yes: $u$ has no critical points in $U$. Indeed, we can have a look at what is possibly (to my knowledge) the most recent paper on these matters, i.e. [1], by Giovanni Alessandrini. He considers the more general elliptic equation (not necessarily in divergence form)
$$
\begin{cases}
\displaystyle\sum_{i,j=1}^2 a_{ij} u_{x_ix_j} + \sum_{i=1}^2 b_i u_{x_i} =0 & \text{on }\Omega \\
\displaystyle u= g & \text{on }\partial{\Omega}
\end{cases}\label{1}\tag{1}
$$
where $\Omega$ is a bounded domain, $a_{ij}=a_{ji}\in C^1(\Omega)$, $b_i\in C^0(\Omega)$, and $g\in C^0(\bar\Omega)$ and proves the following theorem.
Theorem 1.1 ([1], §1, p. 232). Let $u \in W^2_\text{loc}(\Omega)\cap C^0(\bar\Omega)$ be a solution of the Dirichlet problem \eqref{1}. If $g|_{\partial\Omega}$ has $N$ maxima (and $N$ minima), then the interior critical points
of $u$ are finite in number and, denoting by $m_l,\ldots,m_K$ their multiplicities, the following estimate holds
$$
\sum_{i=1}^K m_i\le N-1\label{2}\tag{2}
$$
And as we see, if $N=1$ i.e. $g|_{\partial\Omega}$ has only two critical points $a_\pm\in\partial\Omega$ where it respectively reach its maximum value and its minimum value, then as a consequence of \eqref{2} it must necessarily be $K=0$ i.e. $u$ has no critical points on the interior of $\Omega$.
Addendum: some further notes.
Magnanini in his survey [A2] seems to confirm my statement at the begin of the begin of the answer, i.e. that despite being published in 1987, [1] is still the state of the art study on the analysis of critical points of solutions to the Dirichlet problem for general linear elliptic operators.
On the other and, for the equation
$$
\operatorname{div}(\sigma(x)\nabla u)=0,
$$
in their report [A1] Alberti, Bal and Di Cristo show that if the space dimension is bigger than $3$ the knowledge of the maximum at the boundary alone is not sufficient to conclude something about on the structure of the critical points at the interior. In their own words
We show that the situation is different in dimension $n \ge 3$. More precisely, we obtain that for any fixed (Dirichlet or Neumann) boundary condition for $u$ on $\partial X$, there exists an open set of smooth coefficients $\sigma(x)$ such that $\nabla u$ vanishes at least at one point in $X$.
Reference
[1] Giovanni Alessandrini, "Critical points of solutions of elliptic equations in two variables", Ann. Sc. Norm. Super. Pisa, Cl. Sci., IV. Ser. 14, No. 2, 229-256 (1987), MR939628, Zbl 0649.35026.
Addendum references
[A1] Giovanni S. Alberti and Guillaume Bal and Michele Di Cristo, "Critical Points for Elliptic Equations with Prescribed Boundary Conditions", Seminar for Applied Mathematics, ETH Zürich,
Research Report No. 2016-50}, Switzerland, (November 2016), published also on the Archive for Rational Mechanics and Analysis (MR3686001, Zbl 1378.35099).
[A2] Rolando Magnanini, "An introduction to the study of critical points of solutions of elliptic and parabolic equations", Rendiconti dell’Istituto di Matematica dell’Università di Trieste 48, 121-166 (2016), MR3592440, Zbl 1408.35006.
{\rm div}
instead of\operatorname{div}
sets a bad example for those who are learning to code such things by looking at examples here. Notice that $\operatorname{div}f,$ coded as\operatorname{div}f
, has more horizontal space to the right of $\operatorname{div}$ than does $\operatorname{div}(f),$ coded as\operatorname{div}(f)
. You don't get such context-dependent spacing with{\rm div}
, and there are also other differences seen in some contexts. $\endgroup$