Interaction term becomes more significant Model(1):
Y = A + B + $W_1$ + A*B
Model(2):
Y = A + B + $W_{1,2,3,4,5}$ + A*B
A ranges from 0 to 70 while B ranges from 0 to 25; W refer to set of controls.
In model (1) A (p<0.01) and B (p<0.01) is significant but A*B is not significant (p=0.064).
Then after adding various controls that rids the bias off the variable of interest B...
In model (2) A (p<0.01) is still significant but B becomes not significant (p=0.106)
However, A*B is now significant (p=0.003).
Why is this so? I am quite baffled by these findings. How does one explain this phenomenon of an interaction term that is originally not significant becoming more significant thereafter when controls are added?
 A: I suspect that the issue does not really have to do with the controls that you added but rather that one (or both) of your models suffers from multicollinearity. 
Multicollinearity means that it is possible to express one of the features as a linear combination of the other features. In turn, you can predict $Y$ using multiple different models that are all equally good. 
If, for instance, you should out that $A^*B$, $A$ and $B$ were perfectly collinear then you could express $A^*B = A + B$. Thus, you can express Model (1) as
$Y = A + B + W_1 + (A + B) = 2A + 2B + W_1$ 
or 
$Y = (A^*B) + W_1 = W_1 + 2A^*B$. 
In a case like this, you'd be the same kind of weird significance results (basically $A$ and $B$ could be significant, or $A^*B$ could be significant; your program will report whatever the regression converges to). 
There are different ways to identify multicollinearity... though a straightforward approach is to run regressions between sets of features that you believe might be strongly associated, such as:
$A^*B = A + B + W_1$
$A^*B = A + B + W_{1,2,3,4,5}$
You'll be able to identify the issue if you get a really large $R^2$ value for one of them. If this turns out to be the case, then you can usually get more reliable results by standardizing your variables before running your regression.
