Casper’s attack/defense asymmetry is expected to be strong enough that a 51% attack and the resulting unscheduled hard fork will never happen at all. The cost of a single 51% attack on the Ethereum PoS blockchain would be extremely expensive and probably completely ineffective.
A Casper 51% attack is a theoretical attack of the Ethereum PoS blockchain by a cartel composed of a majority of validators. The Casper protocol severely limits a cartel’s ability to earn profits.
If a cartel ever launched a successful 51% attack, then the Ethereum community could coordinate a hard fork and delete the cartel’s deposits. A successful attack might cost $200 million, but the cost of deleting the cartel’s deposits would be relatively small. Within a couple of days of a 51% attack, the Ethereum PoS blockchain would be back in business, and the cartel would suffer a loss of $200 million. The remaining ethholding public would then enjoy a spike in ether value because of a post-attack supply crunch.
If a cartel ever finalized a block “A” and then finalize some competing block “B”, breaking the finality guarantee and creating two incompatible finalized histories, the community can, with relative ease, coordinate a focus on one branch using Casper’s “simple finality reversion”.
There are two mechanisms by which a simple finality reversion could happen:
1. Coordination could take place on social media, through private channels between block explorer providers, businesses, and exchanges, or on online forums. A decision would be made based on the fact that “whichever block was finalized first is the real block”.
2. “Market consensus” is a Casper mechanism whereby both branches are traded on exchanges for a very short period of time until network effects rapidly make one branch much more valuable with the others. Again, “first finalized chain wins”.
Both of these would probably work in tandem to ensure reversion. Once consensus is reestablished, validators and light and full node operators would be able to manually insert the winning block hash into their client software through a special option in the interface so that their node ignores all other chains.
In the end, a 51% attack would be so costly and pointless that it is unlikely that one will ever be launched against Casper.