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.
There have been numerous scandals over the years regarding their lack of care when handling personally identifiable data.
Meta is not who I look to for being extremely diligent with data.