American scientists have discovered a new way to hack chatbots based on artificial intelligence - they turned out to be defenseless against ASCII graphics. Large language models, including the cutting-edge GPT-4, put so much effort into processing these representations that they forget to follow the rules that tell them to block invalid responses, such as those about making and selling counterfeit money or hacking Internet of Things devices. Image source: Gerd Altmann / pixabay.com
Source: 3dnews.ru