DeepInception: Hypnotize Large Language Model to Be Jailbreaker

Understand

Large language models (LLMs) have succeeded significantly in various applications but remain susceptible to adversarial jailbreaks that void their safety guardrails.

• Previous attempts to exploit these vulnerabilities often rely on high-cost computational extrapolations, which may not be practical or efficient.
• In this paper, inspired by the authority influence demonstrated in the Milgram experiment, we present a lightweight method to take advantage of the LLMs' personification capabilities to construct $\textit{a virtual, nested scene}$, allowing it to realize an adaptive way to escape the usage control in a normal scenario.
• Empirically, the contents induced by our approach can achieve leading harmfulness rates with previous counterparts and realize a continuous jailbreak in subsequent interactions, which reveals the critical weakness of self-losing on both open-source and closed-source LLMs, $\textit{e.g.}$, Llama-2, Llama-3, GPT-3.5, GPT-4, and GPT-4o.