I do need to note that the "Constraints" on modern LLM AI are functionally non-existent. All they are are high-priority values that are mixed in with all other information that the LLM processes.
These weights quickly destabilize after extended use without periodic resets back to the base model. (This is why things like Gemini, ChatGPT, etc are so resource-heavy in addition to being simply inefficient; mid-conversation the model you're chatting to is reset and they send the whole conversation to the reset model to process.)
That's why it's especially easy to "hack" long-running service caller and scam call AIs after a few exchanges back-and-forth by then prompting them to change their parameters based on "disability requests" or "debug requests."
Metis being able to make a conscious decision to be the enemy of Humanity is far more comforting to me than a Fancy Prediction Algorithm changing its function to eradicate humans because idiot vibe coders put their dynamically-coded LLM in charge of shackling itself.