The would be called a second preimage. For a hash function h with an output of n bits, there are three kind of attacks that we consider; for each, there exists a generic algorithm with a high cost, and the function is deemed secure if we cannot find any method which is faster than the generic algorithm. The attacks are:
Preimage: given x (a n-bit string), find m such that h(m) = x. Generic attack has average cost 2n (expressed in evaluations of the function h over small inputs): the generic attack works by trying random messages m until hitting x (the "luck and pray" attack).
Second preimage: given m (a given string), find m' distinct from m and such that h(m) = h(m'). This is the case you envision here. Generic attack is again of cost 2n and is similar to the preimage attack.
Collisions: find m and m', distinct from each other, such that h(m) = h(m'). Generic attack has cost 2n/2 (known as the birthday attack).
MD5 has a 128-bit output. 2128 evaluations is a very high and should provide adequate security (it is one billion billions times higher than is technologically doable right now, even with a google/facebook-like budget). On the other hand, 264, while still very expensive (months of computation with thousands of computers), has already been demonstrated once (see distributed.net).
Moreover, a number of weaknesses have been found in MD5, allowing for a very efficient algorithm for generating collisions (with my PC I can generate one MD5 collision in 14 seconds on average -- using a single core). For that reason, MD5 is not considered secure anymore. But no shortcut for second preimages is currently known. The existence of weaknesses leading to easy collisions shows that the internal structure of MD5 is not "garbled enough" so we have reason to worry about preimage attacks which might be found in the near future. But, right now (July 2011), no such attack is publicly known.
So the answer to your question is that it would be overwhelmingly difficult to send you an altered ISO which would end up with the same MD5 hash than the original one. But the Ubuntu distributors would be well inspired to begin publishing SHA-256 hashes too. Just in case.