The first way is that it's subjective; you have to trust the initial source of the data to give you the correct validator set to use layer. This is a vector for censoring some stake.
The second, probably more important way, is that validators in PoS can decide to ignore or slash a particular validators stake. This can't be done in PoW because the hashes are fungible and stake is not.
AFAIK, the PoS in ethereum is structured that it is extremely hard to ignore a stake (you're vulnerable to slashing if you do), and that you are only vulnerable to slashing if you cheat (not if you are merely offline, at which point you just forego any rewards). Counteracting this requires a similar or greater amount than a PoW chain would require for similar effects (where a 51% attack can also simply refuse to build off of a particular miners blocks, thus effectively stopping their reward).
The second, probably more important way, is that validators in PoS can decide to ignore or slash a particular validators stake. This can't be done in PoW because the hashes are fungible and stake is not.