If you merely put 10 LLM's on the outbound traffic log non of them are going to report something strange going on? I'm not buying it.
This might be helpful reading: https://www.lesswrong.com/w/nearest-unblocked-strategy
As AI systems get smarter, we may reach a point where we have to get it right on the first try or face truly catastrophic consequences: https://www.youtube.com/watch?v=7wy3xyoXYt8