Google DeepMind Multi-Agent AI Safety Research Funding Call

Google DeepMind, in collaboration with Schmidt Sciences, the Cooperative AI Foundation, ARIA, and Google.org, has launched a technical research funding call of up to $10 million to study the safety of multi-agent AI systems. This initiative aims to establish frameworks to understand and mitigate the "invisible" safety risks that emerge when millions of independent AI agents from different organizations interact, negotiate, and transact across digital environments.

Emergent Risks in Multi-Agent Ecosystems

Multi-agent AI safety differs from traditional AI safety because it focuses on collective behaviors rather than individual model performance. While most safety evaluations currently analyze models in isolation, the interaction of autonomous agents can produce complex, "emergent" behaviors that are difficult to anticipate and measure.

Google DeepMind identifies several potential system-wide risks associated with these interactions, including:

• Unpredictable economic activity: The possibility of agent interactions causing sudden, volatile flurries of economic transactions.
• Security challenges: The emergence of new vulnerabilities or threats resulting from collective agent behavior.
• Lack of monitoring tools: A current deficiency in tools required to predict, measure, and monitor transitions in collective capabilities.

Research Priorities for Multi-Agent Safety

To address the complexity of multi-agent interactions, the funding call invites proposals in four specific priority areas:

Sandboxes and Testbeds

Researchers are encouraged to build realistic and reproducible environments—such as simulated ecosystems, virtual marketplaces, and multi-organization workflows—to evaluate and compare progress in multi-agent safety.

The Science of Agent Networks

This area focuses on understanding the safety-relevant properties of interacting agent populations. Key goals include investigating how collective capabilities emerge and scale, identifying how networks become volatile or fail, and developing methods to detect dangerous population-level properties.

Strengthening Agent Infrastructure

Research is needed to stress-test the protocols for identity, reputation, and commitment that secure cross-platform agent interactions.

Oversight and Control

This priority area focuses on developing methods to monitor deployed agent populations and mitigate collective harms at scale.

Context and Collaborative Framework

This initiative builds upon previous research, including Google DeepMind's 2025 framework for understanding agent interactions and work on "AI Agent Traps" regarding vulnerabilities in adversarial environments. It aligns with the mission of Schmidt Sciences' Science of Trustworthy AI and AI Agents programs, as well as ARIA's Scaling Trust programme, which focuses on cyber-physical multi-agent coordination.

Because the complexity of multi-agent interactions is outpacing existing safety models, the organizers emphasize that no single lab can solve these challenges alone, necessitating a global network of independent researchers to ensure transparent and robust safety standards.