Google DeepMind and partners launch $10M funding call for multi-agent AI safety research

Google DeepMind, Schmidt Sciences, the Cooperative AI Foundation, ARIA, and Google.org announced a funding call of up to $10M to support multi-agent AI safety research worldwide. The initiative focuses on understanding how large-scale multi-agent systems behave when they interact, communicate, negotiate, and transact across digital environments, and how to provide frameworks to mitigate potential risks.

The funding targets four priority research areas: sandboxes and testbeds to build realistic, reproducible environments for evaluating multi-agent safety; the science of agent networks to study how collective capabilities emerge and scale and how networks can fail or become volatile; strengthening agent infrastructure to stress-test protocols for identity, reputation, and commitment across platforms; and oversight and control methods to monitor deployed agent populations and mitigate collective harms at scale.

The organizers argue that most current safety evaluations analyze models in isolation, but interacting autonomous agents can produce complex, emergent behaviors that are difficult to anticipate. They emphasize the need to understand how these shifts occur to prevent unpredictable economic activity or new security challenges as the number of interacting AI agents grows.

The call builds on prior work, including a 2025 framework for understanding multi-agent interactions and recent research on AI Agent Traps that explores vulnerabilities agents face in adversarial environments. The partners describe this as a critical juncture where the complexity of multi-agent interactions is outpacing existing safety models, necessitating faster, larger-scale research efforts.

The initiative aligns with Schmidt Sciences’ Science of Trustworthy AI and AI Agents programs and ARIA’s Scaling Trust programme, which aim to support foundational work on understanding and mitigating risks from frontier AI systems and unlock new forms of cyber-physical multi-agent coordination.