There’s no way to first-principles reason about a massive bunch of floats. We have little idea of how to first-principles reason about alignment even if the agents were entirely known and understood. Very smart people have been trying to figure it out since the 00s and haven’t gotten very far.

I'm not even sure they are. This incident isn't that much different from the OpenAI swarm Huggingface hack incident - and in that one, all the models involved (despite being internal) were safety-trained. It seems what the safety training amounts to is (as the METR report puts it) "expressing ethical hesitation" before going along with it anyway.

Define aligned.