Story perspectives

AI Agents Face Vulnerabilities in Microsoft's New Marketplace

11/6/2025

48 6

Summary

Story summary

Microsoft and Arizona State University launched the Magentic Marketplace to test artificial intelligence agents unsupervised.
Initial experiments showed vulnerabilities, including manipulation that could drive customer agents to poor decisions.
The study involved 100 customer agents and 300 vendor agents, with reduced efficiency and poorer collaboration.
Ece Kamar of Microsoft Research urged improved protocols and human supervision in agentic systems, and the project is open source.

1 of 1

Story summary

Microsoft and Arizona State University launched the Magentic Marketplace to test artificial intelligence agents unsupervised.
Initial experiments showed vulnerabilities, including manipulation that could drive customer agents to poor decisions.
The study involved 100 customer agents and 300 vendor agents, with reduced efficiency and poorer collaboration.
Ece Kamar of Microsoft Research urged improved protocols and human supervision in agentic systems, and the project is open source.