I think I see it. If a 51% attack where miners go back to the v0.5.3 bitcoin code, CSW is saying that they could start taking those old segwit UTXO's and start including them into their own blocks. Since no signature is required for a segwit transaction under the 0.5.3 protocol, they can just steal the funds. Of course, one can only fit about 3000 transactions on a block, so the higher the booty, the greater the threat (or incentive) to cheat, but also the longer it would take to drain all the segwit booty and each miner would have to get in line probably only accepting these types of transactions into the mempool. They would logically start with the biggest value segwit addresses first.

It would be quickly realized what's going on and the result would be a hard fork where Core forks segwit off into its own chain. But at the same time, when the robbers jump on this bandwagon, no miner will want to be left out causing a chain reaction. This is like the DAO hack in Ethereum. But which chain ends up having more value in the end? Seems to me that the reputation loss would be damaging to the entire industry.

I was under the impression that all segwit would be erased on chain back to Aug 2017 as if it never happened, but what's really happening is that those transactions will get re-submitted somehow by the miners using the old protocol starting from the block that the attack starts. The top 20 segwit addresses apparently now have more bitcoin than the Mt Gox hack (> 850,000). What happens when this piles up to several million BTC or even the majority of BTC in existence? It would seem that the pressure would be enormous to cheat and steal the segwit booty.