The nature of the 51% attack is often obscured so I would like to expound upon it generally (for all systems).
If someone has 51% of the bitcoin hashing power they can exclude blocks produced by the other 49% and thus end up producing 100% of the blocks actually included in the blockchain. Difficulty will adjust down by 49% and block rate will return to 10 minutes (after being 20 minutes). The other miners (in the 49%) would have to get on the bandwagon and ratify the blocks produced by the attacker (and not produce blocks that violate the attackers rules).
So someone with 51% of the blocks has the power to enforce a STRICTER set of transaction rules. This allows accounts to be frozen, etc.
Likewise with Nxt, if 51% of the active forgers (which is less than 51% of the shareholders) decide to collude they can exclude the 49% of the active forgers, those 49% will be 'punished' for not producing a block and the 51% will eventually control 100% of the active forging. Once again they gain the ability to enforce a STRICTER set of transaction rules and this allows frozen accounts.
Likewise with Ripple, if 51% of the original UNL decide to exclude someone from their UNL then their opinion is ignored.
Likewise with DPOS if 51% of the delegates collude, they can ignore blocks produced by the 49%.
The principle here is that consensus always has a tipping point when the 'majority' goes with it. You cannot enforce a consensus of 90% without the potential of deadlock. These systems must allow people to abstain to be robust against failure because if you require everyone to reach 90% agreement then if 11% of the network goes down you have no consensus.
Recognizing this reality, we can embrace it fully and allow 51% of the delegates to fire any other delegate. It wouldn't actually change the security model, it would simply make it slightly easier to coordinate among the delegates and the rest of the network.
All networks are only as secure as the trust you place in the 51%.