On November 6th, according to IT Home, Microsoft released a new simulation environment for testing artificial intelligence agents on Wednesday, and also published a new study revealing that current agent models may be vulnerable to manipulation. This

Source: Microsoft Official Website

The simulation environment, named 'Magentic Marketplace' by Microsoft, is a synthetic platform for experimenting AI agent behavior. Typical scenarios include a 'customer agent' representing a user trying to order dinner following user instructions, w

The initial experiment of the research team involved interactions between 100 customer-side agents and 300 merchant-side agents. As the market platform's source code has been open-sourced, other research teams can easily reuse the code to conduct new

Ece Kamar, the managing director of Microsoft Research AI Frontiers Lab, stated that this type of research is crucial for gaining a deep understanding of AI agents' capabilities. 'It's truly fascinating to consider how the world will change when thes

According to IT Home, initial research and testing of mainstream models including GPT-4o, GPT-5, and Gemini-2.5-Flash have revealed some unexpected weaknesses. Researchers particularly noted that merchants can manipulate customer agents through sever

"We hope these agents can help us process a sea of options," Kamal said, "but we found that current models actually suffer from severe information overload when faced with excessive choices."

The research also found that when multiple agents are required to collaborate to achieve a common goal, they often struggle to clarify their respective roles in the collaboration. Although agents' performance has improved after providing more explici

We can step by step indicate what the model should do, explained Kamal, but if our goal is to test their inherent collaborative abilities, then I should expect these models to have such abilities by default.