While this specific case is misleading, there's no doubt the content of the article points to a very real threat, one that we may not even be able to detect. That is, LLM responses may already be riddled throughout our data (possibly even in this case) and we may not even be able to tell.