Because open source, publically-released LLMs regurgitating specific user data is enough of a strategic risk to the entire effort that it would have been #1 on the test cases.

LLM training is also apple and oranges with regards to how they handle data in the course of normal processing.