[Edit:]

Edit: Modified attack: https://bitcointalk.org/index.php?topic=2115635.msg21171230#msg21171230, Image: graphical version

Lets say i deposit 1000 BTC to an exchange that accepts the deposit after 6 confirmations. I immediately convert and withdraw in another currency. I am totally free to create an alternative transaction chain starting from the original block with 7 transactions that each pay a high fee (lets say 50 BTC) and that can not be mined in a single block, but only in a chain of 7 consecutive blocks (using nLockTime). I then publish this transaction chain to all miners. Even considering the loss of rewards from the original 6 blocks, the expected reward for the miners is higher, when they discard the original chain and start mining a new chain with my transactions.

Obviously my transactions will not be accepted by regular nodes, but i could communicate them to the miners directly.

If a very small miner got one of the original 6 blocks, his expected reward after switching might be less than what he already got. But i can wait for a situation where only large pools mined the original 6 blocks, that have a large enough probability to profit on my excessive fees.

Miners might fight for the blocks that contain my excessive fees. But they can come up with a solution where they share some of the rewards and it is always more profitable to extend the attacking chain.

The exchange might start to offer a fee reward on the original chain. But every reward it pays is a loss. And i am free to pay the full 1000 BTC. In the end I gained nothing, but forced a transfer from the exchange to the miners. They might want to start this kind of attack.

So, how is it save to accept a large transaction (a block that moved thousands of BTC) as final, before the same amount (some fraction?) was paid in miner rewards? Every moved coin is possibly available to bribe miners to construct an alternative chain.