Somehow, this has me thinking of the Hubris operating system. If a subprogram calls an API incorrectly (wrong number of arguments, wrong argument type, out-of-bounds memory access…), the system kills the caller with no chance for recovery.
If the AI is not supposed to access a thing, and it tries to in a way you know how to detect, halt it and eject it from memory?
If the AI is not supposed to access a thing, and it tries to in a way you know how to detect, halt it and eject it from memory?