I managed to reverse engineer the encryption (refered to as “Obfuscation” in the framework) responsible for managing the safety filters of Apple Intelligence models. I have extracted them into a repository. I encourage you to take a look around.
I find it funny that AGI is supposed to be right around the corner, while these supposedly super smart LLMs still need to get their outputs filtered by regexes.
I feel that if teenagers are able to trivially bypass illegal-word filters by substituting with words that obviously mean the same thing, I think an AGI wouldn't be too inhibited by this either