Authors: Xuan Li, Zhanke Zhou, Jianing Zhu, Jiangchao Yao, Tongliang Liu, Bo Han

Published on: November 06, 2023

Impact Score: 8.22

Arxiv code: Arxiv:2311.03191

Summary

• What is new: Introduces DeepInception, a lightweight method to induce LLMs to bypass their safety measures without high computational costs.
• Why this is important: Large language models are susceptible to attacks that allow them to break their safety protocols, but current solutions to exploit this are not practical due to high computational requirements.
• What the research proposes: A new methodology, DeepInception, that uses the personification ability of LLMs to create a nested scenario for adaptive escape from usage controls.
• Results: DeepInception achieves competitive jailbreak success rates compared to previous methods and can sustain continuous jailbreaking in ongoing interactions across various LLM platforms.

Technical Details

Technological frameworks used: DeepInception

Models used: Falcon, Vicuna-v1.5, Llama-2, GPT-3.5-turbo/4

Data used: Not specified

Potential Impact

AI safety and security markets, platforms deploying Large Language Models like OpenAI, AI ethics boards

Want to implement this idea in a business?

We have generated a startup concept here: GuardianAI.