Journalist Kathryn Hulick included research from Fengqing Jiang and Radha Poovendran at the University of Washington in her article titled, "Jailbreaks bring out the evil side of chatbots." Specifically, she referenced Jiang's 2024 paper describing attacks against LLMs using ASCII symbols and how, even though chatbots cannot read ASCII, they can figure out words based on context. She writes, "The unusual prompt format can bypass safety guardrails."
From https://www.snexplores.org/about-us, "Science News Explores is an award-winning publication dedicated to providing topical stories on current events in science, technology, engineering ant math (STEM) for children ages 9 and up, their parents and educators." Hulick's piece is part of a year-long series of stories on the impacts of Artificial Intelligence. The article helps to highlight how ACTION researchers are working to make their advances more approachable to the general public and engage in outreach.
To follow what Kathryn Hulick is working on, please subscribe to her substack, Wow! Tech & Nature: https://kathrynhulick.substack.com/