The proof is in the paper on Lie algebra cohomology by Chevalley and Eilenberg, in $1948$ (and in Koszul's paper of $1950)$. Chevalley and Eilenberg proved that $H^3(\mathfrak{g},K)$ is nonzero for every nonzero semisimple Lie algebra $\mathfrak{g}$ and every field $K$ of characteristic zero. This works as follows:
Let $B(x,y)$ denote the Killing form of $\mathfrak{g}$. Define a $3$-cochain in $C^3(\mathfrak{g},K)$
by
$$
g(x,y,z):=B(x,[y,z])=B([x,y],z).
$$
In the second equation we use that $B$ is invariant. By Lemma $4.4.4$ of the thesis of Florian Kickinger, page $150$,
every invariant $p$-cochain in $C^p(\mathfrak{g},K)$ of a semisimple Lie algebra over a field of characteristic
zero is a $p$-cocycle in $Z^p(\mathfrak{g},K)$.
Here $g$ is invariant, if
$$
g([x_1,x],x_2,\ldots ,x_n)+\cdots +g(x_1,\cdots,x_{p-1},[x_p,x])=0
$$
for all $x,x_1,\ldots ,x_p\in \mathfrak{g}$. Now, for $p=3$ and the above $g$ we have
\begin{align*}
g([x_1,x],x_2,x_3) & + g(x_1,[x_2,x],x_3)+g(x_1,x_2,[x_3,x]) \\
& = B([[x_1,x],x_2], x_3)+B([x_1,[x_2,x]],x_3)+B([x_1,x_2],[x_3,x]) \\
& = B(([x_1,x_2],x],x_3)-B([x_1,x_2],[x,x_3]) \\
& = 0.
\end{align*}
Hence $g \in Z^3(\mathfrak{g},K)$. Recall that by definition we have
\begin{align*}
0 & = (d_3g)(x,x_1,x_2,x_3) \\
& = -g([x,x_1],x_2,x_3)+g([x,x_2],x_1,x_3) -g([x,x_3],x_1,x_2)\\
& -g([x_1,x_2],x,x_3)+g([x_1,x_3],x,x_2)-g([x_2,x_3],x,x_1)
\end{align*}
By the invariance the first three terms sum up to zero. Hence also the last three terms sum up to zero, i.e.,
we obtain
\begin{align}
0 & = -B([[x_1,x_2],x],x_3)-B([[x_1,x_3],x],x_2)+B([[x_2,x_3],x],x_1).
\end{align}
We claim that $g$ is not a $3$-coboundary. Assume that $g=d_2f$, and recall that
\begin{align*}
(d_2f)(x,y,z) & = -f([x,y],z)+f([x,z],y)-f([y,z],x).
\end{align*}
Now we use the fact, that there is a linear map $T\colon \mathfrak{g}\rightarrow \mathfrak{g}$ such that
$$
f(x,y)=B(T(x),y)=-B(x,T(y)).
$$
Then the identity $g(x,y,z)=(d_2f)(x,y,z)$ gives
\begin{align*}
B(x,[y,z]) & = -B(T([x,y]),z)+B(T([x,z]),y)-B(T([y,z]),x) \\
& = B([x,y],T(z))-B([x,z],T(y))-B(x,T([y,z])) \\
& = B(x,[y,T(z)])-B(x,[z,T(y)])-B(x,T([y,z])).
\end{align*}
Since $B$ is non-degenerate, this implies that
$$
[y,z]=[y,T(z)]-[z,T(y)]-T([y,z])
$$
for all $y,z\in \mathfrak{g}$. This is equivalent to
$$
[ad(z),T]=ad(z)-ad (T(z)) \in ad(\mathfrak{g})
$$
for all $z\in \mathfrak{g}$. Hence there exists a $t\in \mathfrak{g}$ such that
$$
[ad(z),T]=[ad(z),ad(t)]=ad([z,t])
$$
for all $z\in \mathfrak{g}$. Since $ad$ is faithful it follows that
$$
T(z)=[z,t]-z
$$
for all $z\in \mathfrak{g}$. Hence we have
$$
f(x,y)=B(T(x),y)=B([x,t],y])-B(x,y).
$$
Evaluating again $g=d_2f$ and using $(1)$ yields
\begin{align*}
B(x,[y,z]) & = B([x,y],z)-B([[x,y],t],z)+B([[x,z],t],y) \\
& - B([x,z],y)-B([y,z],t],x)+B([y,z],x) \\
& = B([x,y],z) - B([x,z],y) +B([y,z],x) \\
& = B(x,[y,z])+B(x,[y,z])+B(x,[y,z]).
\end{align*}
This gives $2B(x,[y,z])=0$ for all $x,y,z\in \mathfrak{g}$. Since $2\neq 0$ and $[\mathfrak{g},\mathfrak{g}]=\mathfrak{g}$ we obtain that
$B$ is identically zero. This is a contradiction, since $B$ is non-degenerate and $\mathfrak{g}\neq\{0\}$.