In a time-dependent perturbation theory problem, the Hamiltonian is given $H = H_0 + V$, where $V$ is a perturbation that varies sinusoidally with time. $V = V_0 \sin \omega t$. Supposed that the original Hamiltonian $H_0$ is solved, and the (non-degenerate) eigenstates are given by
$$H_0 | n \rangle = E_n |n \rangle$$
Suppose the system is initially prepared in the state $|m \rangle$. The perturbation is switched on for a duration of $T$. I would like to consider the transition probability $P_{m \to n}$ after the perturbation is switched off. If the driving frequency is close to the resonance frequency, one term in the perturbation can be ignored (rotating wave approximation). If I did not make any mistake, the result using first-order perturbation theory is
$$P_{m \to n} = \left|\frac{\langle m |V_0| n \rangle}{E_n - E_m - \hbar \omega} \right|^2 \sin^2 \left( \frac{E_n - E_m - \hbar \omega}{2 \hbar} T \right)$$
This means that the maximum transition probability is proportional to $\left(\frac{1}{E_n - E_m - \hbar \omega}\right)^2$. This quantity can be made arbitrarily large by using driving frequency close to resonance. But the $\sin^2$ term is small, so the small-angle approximation can be used.
\begin{align} P_{m \to n} &\approx \left|\frac{\langle m |V_0| n \rangle}{E_n - E_m - \hbar \omega} \right|^2 \left( \frac{E_n - E_m - \hbar \omega}{2 \hbar} T \right)^2 \\ &= \frac{|\langle m |V_0| n \rangle|^2}{4\hbar^2} T^2 \end{align}
So it suggests that the transition probability is larger when the system is exposed to the perturbation for a longer time. However, the probability should never exceed one. So where did the perturbation theory fail? Or did I make a mistake somewhere?