(This deity is called the stock market)
If such an AI can be reliably made to never ever come back to Earth, they were never a threat in the first place. Nobody knows how to fully test an AI's utility function yet, only randomly test inputs and hope the random distribution we chose is helpful; but every time a diffusion model's output is body horror, every time an LLM makes buggy code (and even every time it gets the pelican-on-bike wrong), this is an example of the test distribution not being good enough.