Egreg's argument is probably the simplest, followed by using cyclicity of the multiplicative group as outlined by Tsemo and Jack. The following approach might be closer to your first idea - utilizing the fact that a finite field has no biquadratic extensions.
So my answer needs the fact that a finite field $F$ has a unique quadratic extension (up to an $F$-isomorphism).
We are given two elements $a,b\in F$ that don't have a square root in $F$. Therefore both extension fields $F(\sqrt a)$ and $F(\sqrt b)$ are quadratic, and hence isomorphic over $F$. This means that we can find a square root of $b$
inside $F(\sqrt a)$. More precisely, it implies that there exist elements $c_1,c_2\in F$ such that
$$
(c_1+c_2\sqrt a)^2=b.\qquad(*)
$$
Expanding the left hand side gives
$$
(c_1^2+ac_2^2)+2c_1c_2\sqrt a=b.
$$
Because $1,\sqrt a$ is a basis for the extension, and the right hand side is an element of $F$ we can conclude that we must have $2c_1c_2=0$. There are three
ways how this could happen:
- If $2=0$, then we are in characteristic two, squaring is injective (see Jack's answer, or infer from Egreg's answer that squaring has trivial kernel), and therefore all the elements of $F$ are squares.
- If $c_2=0$, then $(*)$ is saying that $c_1^2=b$ contradicting the assumption that $b$ has no square root in $F$.
- That leaves the possibility $c_1=0$. So there exists an element $c_2\in F$ such that $ac_2^2=b$. This implies that
$$(ac_2)^2=a(ac_2^2)=ab$$
and therefore in this case $ab$ is the square of an element of $F$.