Interesting to read what @mach1nelearning noticed when he independently replicated the same setup and found LLMs acted in a similar protective way even when a non-AI was the one at risk. (Turns out the original research hadn't tested for what happens when it's a human at risk