"Researchers Reveal 'Deceptive Delight' Method to Jailbreak AI Models"

Palo Alto Networks' Unit 42 researchers have revealed a new adversarial technique they call "Deceptive Delight" that can jailbreak Large Language Models (LLMs) during an interactive conversation by sneaking in a malicious instruction between harmless ones. The simple yet effective method achieves an average 64.6 percent Attack Success Rate (ASR) in three interaction turns. This article continues to discuss observations regarding the Deceptive Delight multi-turn technique.

THN reports "Researchers Reveal 'Deceptive Delight' Method to Jailbreak AI Models"