Language learning models are all about identifying patterns in how humans use words and copying them. Thing is that's also how people tend to do things a lot of the time. If you give the LLM enough tertiary data it may be capable of 'accidentally' (read: randomly) outputting things you don't want people to see.