The ones being used to power "agents" (basically macros) that have been responsible for attacks like HuggingFace.
The LLM needs to be put into a control loop where it activates macros and does actions in "any application on a computer" for this to become a problem.