To prove a protocol is secure in semi-honest model, we have to prove: the view of each party, on each possible pair of inputs, can be efficiently simulated based solely on its own input and and output.
So let's consider this example, Alice holds $x$ and Bob holds $y$, they wants to compute $f(x, y)$ without revealing their inputs. I constructs a protocol which is insecure: Bob simply sends $y$ to Alice, and Alice outputs $f(x,y)$.
But it seems it can be proved according to the definition of semi-honext model. Let's assume Bob is honest, we construct a simulator for Alice. The simulator feeds Bob with a random $y'$, Alice cannot distinguish $y$ and $y'$. So Alice's view can be simulatable. But the protocol is of course not secure.