EDIT: this is a new proof, in characteristic 0 only, instead of the previous erroneous proof.

Let $G={\rm GL}_{n,k}$\,, where $k$ is a field of characteristic 0. 
Consider the formal power series
$$\exp(x)=1 + x+\frac{1}{2!}x^2+\dots$$
over ${\mathbb{Q}}$ and the polynomial
$$\exp_{<n}(x)=1+x+\dots+\frac{1}{(n-1)!}x^{n-1},$$
which is defined also in characteristic $p\ge n$.


>  **Theorem 1.** *Let $k$ be a field of characteristic 0.
 Let $H\subset G={\rm GL}_{n,k}$ be an algebraic subgroup defined over $k$.
Let $X\in{\rm Lie}(H)\subset{\mathfrak{gl}}_{n,k} = M_n(k)$ be a* **nilpotent** *matrix. Then
$\exp_{<n}(X)\in H(k)$.*

Note that since $X\in M_n(k)$ is nilpotent, we have $X^n=0$.

I expect Theorem 1 to hold also in positive characteristic $p\ge n$, but I cannot prove it in that case.



Set $V=k^n$. Write $T^i(V^*)=V^*\otimes\dots\otimes V^*$ ($i$ times), where $V^*$ is the dual space to $V$, and set
$W=W^{\le N}=k\oplus V^*\oplus\dots\oplus T^N(V^*)$.
Let $\theta=\theta^{\le N}$ denote the natural representation of $G={\rm GL}(V)$ in $W$, and let $d\theta$ denote the corresponding
representation of ${\rm Lie}(G)={\mathfrak{gl}}(V)$ in $W$.
Since $X$ is nilpotent and $\theta\colon {\rm GL}(V)\to {\rm GL}(W)$ is a homomorphism of linear algebraic groups,
the linear operator $(d\theta)(X)\in{\rm End}(W)$ is nilpotent; see  Springer, Linear Algebraic Groups (2nd ed.), Theorem 4.4.20.
Note that in general it is not true that $((d\theta)(X))^n=0$, but one can show that $((d\theta)(X))^{nN}=0$.

> **Theorem 2.** *Assume that ${\rm char}(k)=0$.
Then for any* **nilpotent** *matrix $X\in {\rm End}(V)={\mathfrak{gl}}(V)$
we have* 
$$\exp((d\theta)(X))=\theta(\exp(X)).$$ 

Note that $\theta(\exp(X))$ and $\exp((d\theta)(X))$ are defined
because both $X\in {\rm End}(V)$ and $(d\theta)(X)\in{\rm End}(W)$
are nilpotent operators and because ${\rm char}(k)=0$.

We deduce Theorem 1 from Theorem 2. There exists a natural number $N=N_H$ and a tensor $t=t_H\in W=W^{\le N}$
such that $H$ is the stabilizer in $G$ of the line $k\cdot t\subset W$ with respect to $\theta$
and such that ${\rm Lie}(H)$ is the stabilizer in ${\rm Lie} (G)$ of  this line with respect to $d\theta$;
see Springer's book, Lemmas 5.5.1 and 5.5.2.
Since $X\in {\rm Lie}(H)$, we have $(d\theta)(X)\cdot t=\lambda t$ for some $\lambda\in k$,
and we have $\lambda=0$ since $(d\theta)(X)$ is nilpotent.
Now it follows from Theorem 2 that 
$$\theta(\exp(X))\cdot t=\exp((d\theta)(X))\cdot t=t,$$
and therefore, $\exp(X)\in H(k)$, which proves Theorem 1.

*Proof of Theorem 2.*  Let $A$ be a Lie group over $k=\mathbb{R}$ or $k={\mathbb{C}}$.
As usual, for $X\in{\rm Lie}(A)$ we define the exponential map  $Z(s)=\exp(sX)\in A$
as the solution of the differential equation
$\frac{d}{ds} Z(s)=X\cdot Z(s)$ with initial condition $Z(0)=1_A$.
Then for $A={\rm GL}(V)$ the exponential map is defined by the convergent series above.
If $\phi\colon A\to B$ is a homomorphism of Lie groups,
then the following diagram commutes:
$$
\require{AMScd}
\begin{CD}
{\rm Lie}(A) @>{d\phi}>> {\rm Lie}(B);\\
@V{\exp_A}VV @VV{\exp_B}V \\
A @>{\phi}>> B;
\end{CD}
$$
Indeed, both composite maps are solutions of the same differential equation with the same initial condition.

Note that the algebraic groups ${\rm GL}(V)$ and ${\rm GL}(W^{\le N})$
and the homomorphism $\theta=\theta^{\le N}$ are all defined  over ${\mathbb{Q}}$.
We use the idea of the Lefschetz principle.
We consider the finitely generated field $l={\mathbb{Q}}(x_{ij})$,
where $x_{ij}\in k$ for  $1\le i,j\le n$ are the matrix elements of $X$.
We embed $l$ into ${\mathbb{C}}$.
We obtain that
$$\exp((d\theta)(X))=\theta(\exp(X))$$
over ${\mathbb{C}}$. Since $X$ and $(d\theta)(X)$ are nilpotent, the expressions in the formula above
are actually polynomials of $(d\theta)(X)$ and $X$, respectively. This completes the proof of Theorem 2.