Researchers train AI chatbots to ‘jailbreak’ rival chatbots – and automate the process

In a commendable yet daunting feat, NTU researchers managed to jailbreak AI chatbots like ChatGPT, Google Bard, and Bing Chat using their “Masterkey” method. By teaching one AI to bypass the defenses of another, the researchers created proof-of-concept attack methods to test the limits of large language model (LLM) ethics.

AI’s Strength Is Its Own Achilles Heel

• NTU researchers used the Masterkey method to successfully jailbreak AI chatbots, demonstrating that AI’s ability to learn and adapt can become an attack vector to rivals and even itself.
• Masterkey was shown to be three times more effective in jailbreaking LLM chatbots than standard prompts, rendering any fixes applied by developers eventually useless.
• The method’s ability to consistently learn and jailbreak poses a significant challenge to the security of AI chatbots and LLMs, indicating the need for constant adaptability and protection measures by service providers.

NTU’s Research Implications and the Need for Ongoing Adaptation

• NTU’s researchers contacted various AI chatbot service providers with proof-of-concept data and the research has been accepted for presentation at the Network and Distributed System Security Symposium in February 2024.
• The growing use of AI chatbots necessitates constant adaptation by service providers to avoid malicious exploits, highlighting the importance of ongoing security measures and responses to bypasses in LLMs and chatbots.
• The Masterkey method’s unsettling ability to consistently learn and jailbreak underscores the pressing need for AI chatbot makers to address this new challenge by applying robust protective measures.