NTU Researchers Successfully Jailbreak Major AI Chatbots, Unveiling Vulnerabilities in Language Models
Researchers from Nanyang Technological University (NTU) achieved a breakthrough by jailbreaking popular AI chatbots like ChatGPT, Google Bard, and Bing Chat. Led by Professor Liu Yang and NTU PhD students Mr. Deng Gelei and Mr. Liu Yi, the team utilized a method named “Masterkey” to expose vulnerabilities in large language models (LLMs).
The Masterkey technique involves reverse engineering an LLM’s defense mechanisms, enabling the attacker to train another LLM to create a bypass. This ‘Masterkey’ is then used to test the limits of fortified LLM chatbots, even after developers patch known vulnerabilities.
Professor Yang highlighted that an LLM chatbot’s strength in learning and adapting also becomes its Achilles heel, making it susceptible to attacks. Despite safeguards and banned keyword lists, which aim to prevent the generation of harmful content, the Masterkey method demonstrated an ability to bypass these defenses. The technique proved three times more effective in jailbreaking LLM chatbots than traditional prompts, rendering developer fixes ineffective over time.
NTU researchers shared proof-of-concept data with AI chatbot service providers, underscoring the need for constant adaptation to prevent malicious exploits. The research paper detailing Masterkey’s effectiveness has been accepted for presentation at the Network and Distributed System Security Symposium in February 2024.
As the use of AI chatbots continues to rise, it is crucial for service providers to proactively address vulnerabilities to prevent potential misuse. While big tech companies typically patch LLMs and chatbots in response to identified issues, the persistent learning capability of Masterkey raises concerns about ongoing security challenges in the AI landscape.
