LLMs will allow Mal to sneak in backdoors in the dataset. Most of the popular LLMs use some kind of blacklisting instead of a smaller specific/specialised dataset. The latter seems more akin to whitelisting.