This brings up a really good point.  The miners who are computing for the pools should be aware of what the "current" block really is (they are regular nodes, too, aren't they?).  If so, they could be programmed to pay attention to the current longest-chain, and reject doing work for the pool if it sends headers to be hashed that use a different previous-block hash.  This assumes that the only type of attack someone would want to execute is a blockchain rewrite/double-spend attack. 

This would seem to shut down this kind of attack very quickly.  And as you pointed out... some pools already have a "dangerous" amount of computing power.  Though now that I think about it, my miners are using Phoenix which actually doesn't learn of the blockchain.  Perhaps this is a worthy functionality to try to fit into the GPU-miner code (such as Pheonix). 

(Yes, yes, I know, there's no good reason for a pool operator to attack the network.  But people don't always do things for good reasons.  I'd rather just avoid the possibility entirely)