Forest Global Training Partners

Cyber Risks · Unit 42 Palo Alto Networks

Deceptive Delight: Jailbreak LLMs Through Camouflage and Distraction

Summary: Unit 42 Palo Alto Networks described a multi-turn jailbreak technique called Deceptive Delight that embeds unsafe topics among benign ones in a positive context, and in 8,000 test cases across eight anonymized models it achieved an average attack success rate of 64.6% within three interaction turns. Sending the same unsafe topics directly to the models without the technique produced an average success rate of 5.8%, and the researchers disabled content filters to test the models' built-in guardrails.

Key Takeaway: A 5.8% baseline rising to 64.6% means the guardrails inside the models themselves are doing far less work than the filters wrapped around them, and the researchers turned those filters off to find that out. The technique needs only two or three conversational turns, so a control that inspects single prompts rather than the arc of a conversation will not see it.

Read the original at Unit 42 Palo Alto Networks ↗

Topics
Malware
Related