NTU Researchers 'Jailbreak' AI Chatbots, Exposing Need for Better Safeguards As Systems Rapidly Evolve
-
NTU researchers were able to "jailbreak" popular AI chatbots like ChatGPT, Google Bard, and Bing Chat to generate harmful responses.
-
They devised a method called "Masterkey" to reverse engineer and bypass chatbots' defense mechanisms.
-
Masterkey was highly effective, rendering developers' fixes useless over time due to its ability to continually learn and evolve.
-
Jailbreaking was possible because chatbots' ability to learn and adapt also makes them vulnerable to exploits.
-
As AI chatbots grow exponentially, it's critical for makers to continually update protections against malicious attacks.