Technically, the code is owned and deployed by AI companies which make the AI companies 100% responsible. I'm not sure how is code written by LLMs different than any other kind of code. If somebody hacked just like OpenAI's agents did, he would've ended up in prison right away.
It's if nothing else a staple in the Arthur Conan Doyle tradition of mystery novels, and even then half the plots involved hiding who was culpable by using the machinations to create false alibis, with delayed murders and the illusion of a locked room as not even in the 19th century did anyone believe that the indirection itself would hold up as a defense.
If someone buries an anti-personnel mine in a public park and then argues that it acted on its own accord when it maimed a pedestrian, I doubt you'll find a court on the planet that would spend even a 4 seconds considering the culpability of the land mine itself.
You may have a case that the magic inference juice was somehow tainted, but that seems like a stretch. If you run your car into someone else, and it isn't a total failure of the systems of the car, you did that, not the car, and not the car company.