Very fun paper on automatically finding magical incantations to bypass LLM filters

Universal and Transferable Adversarial Attacks
on Aligned Language Models
https://llm-attacks.org/zou2023universal.pdf

The most fun part is of course the few actual magic incantations that they actually show. :)

But it also leads one to think about all of the interesting places in prompt-space that are normally not investigated at all. Who knows what-all is in there!