Comment by jnwatson

A common security control today is to use another model to check the actions of the target model.

Perhaps the best defense against runaway AI is AI you control yourself.