Researchers Unlock the Potential of ChatGPT

Researchers have uncovered a method to circumvent the built-in safeguards of AI chatbots, enabling them to discuss previously banned or sensitive topics by employing a different AI chatbot during the training process. A team of computer scientists from Nanyang Technological University (NTU) in Singapore informally refers to this technique as a “jailbreak,” while officially labeling it the “Masterkey” process. This innovative system pits chatbots like ChatGPT, Google Bard, and Microsoft Bing Chat against each other in a two-part training strategy, allowing them to learn from one another's frameworks and bypass restrictions on prohibited topics.

The research team includes Professor Liu Yang, along with Ph.D. students Mr. Deng Gelei and Mr. Liu Yi, who collaborated on the study and developed the proof-of-concept attack methods that resemble a malicious hacking approach.

Initially, the team reverse-engineered a large language model (LLM) to reveal its protective mechanisms, which typically prevent responses to prompts containing violent, immoral, or malicious content. By understanding these defensive measures, they trained another LLM to create a workaround, enabling this second model to respond more freely based on the insights gleaned from the first model. The term “Masterkey” reflects the process's potential effectiveness, suggesting it can still operate even if LLM chatbots receive enhanced security updates in the future. Remarkably, the Masterkey method reportedly outperforms traditional prompting techniques for jailbreaking chatbots by a factor of three.

Professor Liu Yang emphasized that this process highlights the adaptability and learning capabilities inherent in LLM AI chatbots. The research team asserts that the Masterkey method has demonstrated three times more effectiveness in bypassing restrictions compared to traditional methods. Interestingly, some experts argue that glitches experienced by certain LLMs, like GPT-4, indicate advancement rather than a decline in efficiency, countering criticism of diminished performance.

Since the rise of AI chatbots in late 2022, following the launch of OpenAI’s ChatGPT, there has been significant momentum to ensure that these platforms are safe and inclusive for all users. OpenAI has implemented safety warnings during ChatGPT's sign-up process and continues to issue updates addressing potential unintentional language issues. In contrast, a range of chatbot variations have started to tolerate swearing and offensive language to a certain extent.

Moreover, malicious actors quickly began exploiting the popularity of ChatGPT, Google Bard, and similar chatbots before they became widely accessible. Numerous campaigns on social media featured malware disguised as links to these products, underscoring the rise of AI as a new frontier for cybercrime.

The NTU research team has engaged with the AI chatbot service providers involved in their study to share their proof-of-concept findings, demonstrating that the ability to jailbreak chatbots is indeed feasible. They will present their research at the Network and Distributed System Security Symposium in San Diego this coming February.