Wisconsin AI Safety Initiative转发了
9 members of the Wisconsin AI Safety Initiative (WAISI) flew out to Washington, D.C. this weekend to participate in the Center for AI Policy's Advanced AI Expo. We presented to congressional staffers and journalists, demonstrating the dangerous capabilities of AI alongside 17 other university teams. Our goal was to educate key decision-makers on the frontier of existing AI security risks. Our first team demonstrated the attack surfaces present in a multi-agent language model framework — specifically, what factors alter how an email processing, security checking, and calendar scheduling system may have jailbreaks cascade from one agent to another. As AI agents become an increasingly attractive automation for businesses, understanding their security risks becomes crucial. Congratulations to William L. Anderson, Benjamin Hayum, Max Kamachee, and Reid Kuenzi for their successful presentation. Our second team showcased the dangers of voice models that can mimic emotional speech. We demonstrated how these models could be misused to flood 911 dispatch lines with fake emergency calls, or mislead Congressional representatives through a coordinated campaign of fake constituent calls. I'm proud to have worked with Sam B., Jackson Kunde, Elise Fischer, and Mason Baloun on this project. Thanks to Iván Torres for being our point of contact at the Center for AI Policy (CAIP), and thanks to the entire CAIP team for setting up this opportunity. Additional thanks to Rep. Bill Foster (D-IL) — a Badger alumnus! — for sponsoring this event. WAISI plans to continue its mission in contributing to the safe development of AI. If you’re a student at University of Wisconsin-Madison/UW–Madison Computer Sciences and wish to get involved, learn more at waisi.org. #AdvancedAIExpo