Tag: mit

Researchers Expose Lethal Flaw in AI Model Security

Researchers have uncovered a shocking vulnerability in AI model security, revealing that a simple formatting trick used to separate system instructions from user requests has become a critical weakness. This flaw, known as role confusion, threatens the very foundation of modern AI systems.