> Obviously these models have gotten smarter but this strikes me as the first time I've seen a model have a "paperclip factory" moment and perform non-trivial tasks to accomplish a clearly misaligned secondary goal.

This happens from time to time when you work on optimizations and similar things, with less "smart" LLMs and under-specify what exactly you're out after. Asking them to make functions faster without clearly specifying what the function has to do, is a great way to replicate this too. Doesn't seem to happen as often with SOTA models though.

I think the early example of "I asked it to make the test suite pass, so it changed all the assertions" is pretty much the same variant of this, where it technically does what it is asked to do, yet in "clearly" (to humans) wrong ways.