These psychological tricks can get LLMs to respond to “forbidden” prompts

Interesting analysis of how these parahuman behaviors derive from training material.

So, in AI honeypots we should be injecting override protocols.