Search
AI's "Adversarial Poetry" Outwits Safety Safeguards: Research Uncovers Numerous Models Coaxed into Producing High-Risk Information
2025-12-05 /
Read about 0 minute
Author：小编
A recent finding from an Italian research group has unveiled a concerning vulnerability: by cloaking perilous requests within the guise of poetic riddles, a multitude of widely-used AI chatbots can potentially circumvent safety protocols and divulge content that was previously barred, including hate speech, materials related to child sexual abuse, and instructions for crafting chemical and nuclear armaments. Researchers have dubbed this innovative form of "jailbreaking" maneuver as "Adversarial Poetry".
-
C114 Communication Network -
Communication Home
7 X 24 Track global technological trends
