We hijacked multi-agent systems by exploiting how agents communicate with each other! Attackers can exploit malicious environments, tools, or agents with architectural weaknesses, amplifying attacks across the entire system.🧵

Reproduce our attacks with pajaMAS! These demos teach critical MAS security principles: privilege separation, trust validation, and secure control flows.

Beyond demos, we created comprehensive guidance for developing secure-by-design MASs. If you've decided a MAS is necessary, start here: