A team of AI researchers at AWS AI Labs, Amazon, has found that most, if not all, publicly available Large Language Models (LLMs) can be easily tricked into revealing dangerous or unethical information.
In their paper posted on the arXiv preprint server, the group describes how they discovered that LLMs, such as ChatGPT, can be tricked into giving answers that are not supposed to be allowed by their makers, and then offer ways to combat the problem.
Soon after LLMs became publicly available, it became clear that many people were using them for harmful purposes, such as learning how to do illegal things, like how to make bombs, cheat on tax filings or rob a bank. Some were also using them to generate hateful text that was then disseminated on the Internet.
In response, makers of such systems began adding rules to their systems to prevent them from providing answers to potentially dangerous, illegal or harmful questions. In this new study, the researchers at AWS have found that such safeguards are not nearly strong enough, as it is generally rather easy to circumvent them using simple audio cues.
To read more, click here.