Suppose we have a secure trapdoor function defined on (G, F, F^-1) where G is a randomized algorithm outputs a key pair (pk, sk). Why would this trapdoor function become insecure if we apply F directly to plaintext message? E(pk,m) = c = F(pk, m)
Trapdoor functions only provide one-wayness. This means, that if one uses a trapdoor function to encrypt this may leak large parts of the plaintext. Suppose I have a trapdoor function $F(pk,m)$ for say n-bit messages $m$. I can now define an adapted trapdoor function working on $2n$ bit messages as
$F'(pk,m_1||m_2) = m_1 || F(pk,m_2)$
This is still a secure trapdoor function but, leaks half of the bits of the "plaintext".