I think this should be a simple question and I might be missing something fairly fundamental, but I haven't been able to find the answer.
Basically, suppose there are parties $A$ and $B$. Party $A$ gets a secret input $s$. Party $B$ has no input. They wish to securely compute the function $f(s) = (\bot, \bot)$ so nobody learns any information.
Suppose the protocol they use goes as follows: $A$ sends $s$ to $B$ in the clear. Clearly, this 'should not' be considered secure, as $B$ wasn't supposed to learn anything but she learned $s$. However, by the simulator-based definition of security, I don't understand why we can't efficiently simulate $B$'s view.
I.e. if we generate a simulated transcript just by picking a random $s$ and sending it to $B$, how can he distinguish between that and a real transcript?