You're thinking about "alive" as "humanlike" as "subjective experience" as "dangerous". Instead, think of agentic behavior as a certain kind of algorithm. You don't need the human cognitive architecture to execute an input/output loop trying to maximize the value of a certain function over states of reality.
> But it isn't anymore dangerous than any other faulty software that has access to sensitive systems.
Seems to me that can be unboundedly dangerous? Like, I don't see you making an argument here that there's a limit to what kind of dangerous that class entails.