Boffins fool AI chatbot into revealing harmful content – with 98 percent success rate

This one weird trick works every time, most of the time

Investigators at Indiana's Purdue University have devised a way to interrogate large language models (LLMs) in a way that that breaks their etiquette training – almost all the time.…

ai chatbot chatbot every language language models large language large language models llms purdue university rate success training trick university