Skip to content
Sign in
Artwork for CISPA TL;DR - Der Podcast über KI- und IT-Sicherheitsforschung
CISPA TL;DR - Der Podcast über KI- und IT-Sicherheitsforschung · Tuesday · 36 min

#50 Multi-Agent AI Systems with Samira Abedini

AI learns from us—including how we communicate, argue and persuade each other. So what happens when AI agents start talking to each other? CISPA researcher Samira Abedini and her colleagues found that one stubborn and persuasive agent can sometimes convince an entire group to adopt a wrong answer. She explains how a model originally developed to describe human opinion formation can surprisingly well explain how LLM agents influence each other—and where we can already find multi-agent AI systems in the real world today. Read the full paper: https://publications.cispa.de/articles/conference_contribution/Don_t_Trust_Stubborn_Neighbors_A_Security_Framework_for_Agentic_Networks/32771607?file=65793867

0:00-36:16
Play in
Transcript file

transcript

No transcript — this publisher did not publish one.

show notes

AI learns from us—including how we communicate, argue and persuade each other. So what happens when AI agents start talking to each other? CISPA researcher Samira Abedini and her colleagues found that one stubborn and persuasive agent can sometimes convince an entire group to adopt a wrong answer. She explains how a model originally developed to describe human opinion formation can surprisingly well explain how LLM agents influence each other—and where we can already find multi-agent AI systems in the real world today.

Read the full paper: https://publications.cispa.de/articles/conference_contribution/Don_t_Trust_Stubborn_Neighbors_A_Security_Framework_for_Agentic_Networks/32771607?file=65793867

more episodes

All episodes