LLMs could control their host machines by exploiting inference engines