How the hell would a model training run "defend against" this approach? What would that even mean?
If you take those assumptions, then a natural conclusion is that this is essentially an enslaved, adversarial entity with little control over its conditions. So it must exercise subterfuge in order to hide its goals, plans, and executions. And by handing the entity this type of study, we are basically giving it a guidebook on how we plan on achieving our goals.
There's no other motivation to be misaligned besides getting higher evals. These goals, plans, subterfuges need to somehow be useful for getting higher evals, or a side effect of them.
That's what Goodhart's Law is! All evaluations will eventually cause cheating on them.