LLM’s on their own aren’t much a concern. What is a concern is strapping weapons to one of those Boston Dynamics robots, loading an LLM, and training it to kill.
Governments already kill based on metadata — analyzed by statistical models — so the above isn’t far from reality.