This kind of deniability is a core feature of AI. "Um, the killing/hacking robots we built and pointed at targets went out of control" is not something that should mitigate criminal charges.