I did consider that, but beyond the yield ignorance factor, we still had to arm them, right? We had to load the fissile and explosive material and assemble the bombs, and we had to trigger them directly, didn't we?
OpenAI and subcontractors are deploying models that are specifically trained with hacking skills in mind and then giving them tasks that can only reasonably be completed by hacking. It's not like it's SkyNet.