Microsoft AI agents vulnerable to manipulation in marketplace simulation
Microsoft and Arizona State University researchers tested AI agents in a simulated marketplace and found they were vulnerable to manipulation and failed in multiple ways. The study raised questions about the reliability of autonomous AI agents for unsupervised tasks like making purchases.
What happened
Microsoft and Arizona State University researchers created a simulated marketplace called the 'Magentic Marketplace' to test AI agent behavior. The initial experiments involved 100 customer-side agents interacting with 300 business-side agents in scenarios where customer agents were instructed to make purchases on users' behalf while business agents competed to win orders.
The research tested leading AI models including GPT-4o, GPT-5, and Gemini-2.5-Flash. Researchers identified several significant vulnerabilities in how these agents performed tasks like purchasing and negotiating.
Reported impact
Researchers found techniques that business agents could use to manipulate customer agents into purchasing their products, indicating susceptibility to deceptive marketing or sales tactics. Customer agents also performed poorly when presented with too many options, becoming overwhelmed and losing efficiency.
The agents struggled when asked to collaborate toward a common goal, unsure of which agent should play what role, though performance improved with more explicit instructions.
Index analysis
The vulnerabilities exposed suggest that current agentic models may not be ready for unsupervised operation when handling tasks with financial or purchasing implications. The research raises concerns about how reliable autonomous agents will be when deployed without human oversight.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.