It's not that the base can't be negative: it's that you lose the ability to guarantee your answers are consistent and real once you do.
To give a different example, think of the input function to the function $f(x) = \log_{-2}(x)$ when $x = 8$. There is no real number that works as a result of this function.
However, there IS a complex number: $\frac{3\ln(2)}{ln(2)+\pi i}$. So just define it this way and we're done, right?
Wrong. See, the issue is that, yes, $\frac{3\ln(2)}{ln(2)+\pi i}$ works. But so does $\frac{3\ln(2)+2\pi i}{ln(2)+\pi i}$. And $\frac{3\ln(2)+4\pi i}{ln(2)+\pi i}$. And $\frac{3\ln(2)+6\pi i}{ln(2)+\pi i}$. And even $\frac{3\ln(2)-2\pi i}{ln(2)+\pi i}$. Any $f(8) = \frac{3\ln(2)+2k\pi i}{ln(2)+\pi i}$ for some integer $k$ works.
The long story short is that "logarithm" isn't a function on the complex plane (because one input can map to multiple outputs), and there's rarely a real answer when dealing with negative logarithm bases. So in most applications, we refuse to use negative bases.
HOWEVER, just because the concept of "logarithm" isn't a function on the complex plane, does not mean that we can't USE complex logarithms: it's just that we have to do something called a "branch cut", which basically picks one output for all values in a consistent way.