Research
Deceptive Delight: Jailbreak LLMs Through Camouflage and Distraction
First reported · Discovered paloaltonetworks.com
Page published
Earliest dated coverage: 23 Oct 2024 · First observed: 28 Oct 2026 · Latest dated coverage: 23 Oct 2024
Coverage timeline
Single-source research — one report is available.
Why it matters
Multi-turn jailbreaks such as Crescendo show why defenders must evaluate safety across conversation history rather than relying solely on filtering individual prompts.
Unit 42’s “Deceptive Delight” article addresses LLM jailbreaking through camouflage and distraction. The linked Crescendo study demonstrates and evaluates a multi-turn jailbreak that starts with benign questions and progressively references the model’s replies to elicit prohibited outputs across several public model families.