This logic is, as you say, simplistic. Anyone with such a large stake can cash out, and then use their former large stake to fabricate an alternative chain in which they did not cash out, forking the blockchain at the block just before they cashed out. Commonly proposed defenses against this (checkpointing) are limited because when new nodes enter the network, they are in a naive state where they do not know who to trust.
In a proof of work system, the entire blockchain can be validated by a naive node. Fraudulent blockchains could be constructed and fed to naive nodes, but they would be obvious forgeries because they would have to be constructed at lower network difficulties in order to generate blocks faster than the real network and become the longest chain.
In a proof of stake system, people who formerly had stake in the system can simulate fraudulent alternative histories forked at the point before they cashed out, and feed those histories to naive nodes. If enough naive nodes accept the alternative history, it becomes the accepted chain. The possibility exists that these former stakeholders might use some of their cashed out money to start up many nodes in the cloud and feed their alternative chains to those nodes.
The only real way to defend against this is to have trusted nodes coded into the protocol provide bootstrapping data to naive nodes. Once you do that, your network is no longer trustless.
Read more here: http://www.cs.cornell.edu/~iddo/CoAslides.pdf