This is a 51% attack where the attacker sets the block timestamps artificially to lower the difficulty, allowing him to profit more coins than his hash rate would warrant. It works because the difficulty adjustment formula is based purely on timestamps that can be chosen quite freely by the attacker. In particular, the timestamps are not required to be in increasing order. ArtForz explains the attack in detail here.
This attack was used against the new Geist Geld block chain. Geist Geld difficulty adjusts every 16 blocks making the chain particularly vulnerable against this attack. An attacker targeting Bitcoin would have to mine a full set of 2016 blocks at the current difficulty before starting to get any benefit compared to a "normal" 51% attack.
Obviously, performing this attack would be horrendously expensive. (Coinometrics places the cost at about $463 million at time of writing.)
However, if you were successful, you could lower the difficulty and increase the block rate by a factor of 4 every 2 weeks, letting you mint coins more than every 10 minutes.
The obvious way of fixing this - changing the block retarget period from (2016*n, 2016*n + 2015) to (2016*n, 2016*n + 2016) is problematic. Blocks produced by the new implementation would be invalid in the old implementation, and vice-versa. (Source)
So, how could a time warp attack be prevented in a non-disruptive way?