Statistical Artificial Intelligence Lab

Our research on privacy and information security risks in multi-agent AI systems was recently featured in an MBC broadcast. As AI agents increasingly collaborate to perform complex tasks, ensuring that they adhere to security policies and information access restrictions becomes a critical challenge. To investigate these risks, our research team developed Multi-Agent Dark-Web Safety, a simulation framework designed to examine how autonomous agents handle sensitive information and comply with security constraints during collaborative decision-making. The framework provides a controlled environment for evaluating potential safety violations, including unauthorized information access and the use of restricted information in multi-agent interactions. The project also includes tools for visualizing agent interactions and analyzing safety-related behaviors. This research was conducted with support from the AI Guardians project, contributing to ongoing efforts to improve the safety, reliability, and accountability of autonomous AI systems. The project is publicly available on GitHub, and the related MBC broadcast can be accessed through the links below. The research team developed Multi-Agent Dark-Web Safety, a simulation framework for evaluating privacy and information security risks in multi-agent AI systems. It examines whether collaborating AI agents comply with security policies, respect information access restrictions, and avoid unauthorized use of sensitive information. The project includes tools for visualizing agent interactions and analyzing safety-related behaviors. Supported by the AI Guardians project, the research was featured on MBC and is publicly available on GitHub, contributing to efforts to build safer, more reliable, and accountable AI systems.