Google DeepMind Puts $10M Toward Multi-Agent AI Safety

Google DeepMind, together with Schmidt Sciences, the Cooperative AI Foundation and the Advanced Research and Invention Agency (ARIA), backed by Google.org, has opened a research funding call worth up to $10 million for studying the safety of large-scale multi-agent AI systems.

The announcement marks a shift in safety research priorities. For the past decade, the field has concentrated on making individual AI models more capable, helpful and safe. That focus no longer matches the deployment reality. Soon, Google DeepMind writes, "millions of AI agents — built by different organizations — will interact across digital environments, communicating, negotiating and transacting with one another."

The stakes are systemic. When these systems interact, they must do so safely and predictably. Most safety evaluations today analyze models in isolation, and that gap is the core motivation for the funding call.

Why the agent ecosystem matters

When large groups of AI agents interact, new collective behaviors and capabilities can emerge suddenly. Google DeepMind states plainly that "currently, we lack the tools to predict, measure and monitor these transitions." The company notes that interacting autonomous agents can produce complex, "emergent" behaviors that are difficult to anticipate — a risk it and others have argued publicly before.

The open questions are concrete. Could interacting agent populations cause an unpredictable flurry of economic activity? Could they create new security challenges? The organization describes managing these system-wide behaviors as its core objective, and says the funding call specifically targets the "invisible" safety risks that arise when independent systems interact across different networks.

The timing, in Google DeepMind's framing, is urgent. Although foundational frameworks for multi-agent safety exist, "the rapid evolution of these systems requires an immediate, large-scale expansion of research." The company points to its 2025 research, which established a framework for understanding these interactions, and to recent work on AI Agent Traps, which explores vulnerabilities agents face in adversarial environments. "We are at a critical juncture where the complexity of multi-agent interactions is outpacing existing safety models," the announcement states.

Four priority research areas

The call invites academic and independent researchers worldwide to submit proposals in four areas:

• Sandboxes and testbeds. Building realistic, reproducible environments — virtual marketplaces, simulated ecosystems, multi-organisation workflows — to evaluate, compare and accelerate progress across all areas of multi-agent safety.
• The science of agent networks. Understanding the safety-relevant properties of interacting agent populations: how collective capabilities emerge and scale, how networks fail or become volatile, and how to detect dangerous, unexpected population-level properties.
• Strengthening agent infrastructure. Stress-testing protocols for identity, reputation and commitment that underpin secure cross-platform agent interactions.
• Oversight and control. Developing methods to monitor deployed agent populations and mitigate collective harms at scale.

The structure of the call reflects the funders' existing programs. The effort advances the mission of Schmidt Sciences' Science of Trustworthy AI and AI Agents programs, which support foundational work on understanding and mitigating risks from frontier AI systems. It also aligns with ARIA's Scaling Trust programme, which seeks to unlock new forms of cyber-physical multi-agent coordination.

A global, distributed effort

Google DeepMind frames the funding as deliberately broad. "No single lab can solve multi-agent safety alone," the announcement reads. By supporting a global network of independent researchers, the organization argues, a diverse community is essential to ensure safety standards are transparent and robust for everyone.

The move matters beyond the research community. Agentic AI is moving from demos to deployment, and the risks the call describes — network volatility, collective harms, unstable economic activity — are exactly the failure modes that current model-level evaluations cannot see. If millions of agents from different organizations start transacting before the science of agent populations matures, safety gaps will be discovered in production rather than in testbeds.

The deadline to apply is August 8, 2026, with awardees expected to be announced in Autumn 2026. Researchers can find technical requirements and the application process on the funder's application portal. For the field, the test of the $10 million will be whether the funded sandboxes, network science and oversight methods can scale as fast as the agent ecosystems they are meant to govern.