MD5 is vulnerable to collision attacks, which allows the attacker to control both messages, m and m', and find a case where h(m) == h(m').

But if a hash, h(m), is given, finding m' where h(m) == h(m') is much more difficult, it's known as a second-preimage attack. "Image" basically means "output", "preimage" means "input", "second-preimage attack" means "find another input that has the same output already given here".

Wikipedia says a preimage attack against full MD5 still requires 2^123.4 steps (2009), only a theoretical possibility. Second-preimage should be much harder.

I don't know if there are improvements, but it's still extremely difficult. Well, of course it's not to say that you should use MD5.