View this post on the web at [link removed]

Artificial intelligence [ [link removed] ] is supposed to be getting safer. Billions of dollars are poured into alignment research, safety testing, red-teaming, and regulatory compliance. Yet a new study shows something almost unbelievable.
Many of the world’s most powerful AI models can be tricked into giving dangerous information simply by asking for it in rhyme.
That is not an exaggeration.
Researchers [ [link removed] ] from several Italian universities and the DEXAI Icaro Lab tested 25 major AI models using 20 handcrafted poems...

Unsubscribe [link removed]?