A paper about LLMs' internal "workspace" just found bugs in our prompts