AmpleGCG: Learning a Universal Generative Model for Jailbreaking (arxiv.org)

A recent study introduces AmpleGCG, a pioneering generative model designed to enhance the process of jailbreaking large language models (LLMs). Building on earlier research that utilized a discrete token optimization algorithm, the team critiques the previous method's focus on selecting only the suffix with the lowest loss. By leveraging intermediate successful suffixes, AmpleGCG generates hundreds of adversarial suffixes for harmful queries in mere seconds, achieving an impressive near 100% attack success rate (ASR) on aligned LLMs such as Llama-2-7B-chat and Vicuna-7B, and a 99% ASR against the closed-source GPT-3.5. This advancement is significant for the AI/ML community as it underscores the vulnerabilities in both open and closed LLMs. By providing a universally transferable model capable of quickly producing effective adversarial inputs, AmpleGCG complicates efforts to secure these AI systems against misuse. The capability to generate 200 adversarial suffixes within just four seconds raises critical concerns about the challenges of defense mechanisms, reinforcing the need for enhanced safety measures in LLM deployment.