A new study published in the Journal of Plos One has examined the activities of groups called the Red LLM team, which aims to test the limits of large language models (LLM) such as ChatGPT, forcing these systems to provide unauthorized or unexpected responses.
The study, conducted with a deep ierview with four active individuals in the field, shows that this activity is a combination of curiosity, creativity and collective cooperation, and is often done with non -destructive motivations to ideify systems’ weaknesses.
Researchers have classified five penetration techniques io five categories using data theorizing method.
The study, conducted by Nana Inieh of IT University of Copenhagen and colleagues, emphasized the importance of the human -ceered approach to artificial ielligence security.
The people studied, from software engineers to artists, work together in online societies such as Twitter and Discords, and use creative methods such as using alternative languages ​​or imaginative scenarios to circumve chatter restrictions.
The results of this study show that such activities, in addition to helping to develop safer systems, highlight the need to deepen human behaviors in ieracting with advanced technologies and can be a basis for future research in this field.
