Google DeepMind and Four Partners Commit Up to $10 Million to Multi-Agent AI Safety Research

Google DeepMind announced on June 11, 2026 that it will provide up to $10 million in funding for researchers worldwide studying the safety of interactions among multiple AI agents¹. The program is a joint effort with Schmidt Sciences, the Cooperative AI Foundation, the UK’s Advanced Research and Invention Agency (ARIA), and Google.org. Applications close on August 8, 2026, with awardees expected to be announced in Autumn 2026¹.

Research funding focused not on the safety of a single AI model but on interactions between agents is unusual. Rohin Shah, who leads AGI safety and alignment at DeepMind, told MIT Technology Review that “the main issue is that there just isn’t really a field of research for multi-agent safety yet”², suggesting the funding aims to bootstrap the research field itself.

Why Interactions Between Agents Are Becoming a Problem

DeepMind envisions an era in which millions of AI agents, built by different organizations, interact across digital environments — communicating, negotiating, and transacting¹. The announcement notes that “interacting autonomous agents can produce complex, ‘emergent’ behaviors that are difficult to anticipate” and that “the complexity of multi-agent interactions is outpacing existing safety models”¹.

Even if each individual agent behaves safely on its own, new collective behaviors and capabilities can suddenly emerge when many agents are combined. According to DeepMind, tools to predict, measure, and monitor these transitions are currently lacking¹.

Shah says agent deployment introduces a whole new class of risk, and expects agents to roll out across the economy in the coming months².

Four Priority Research Areas

The program targets four research areas¹.

The first is sandboxes and testbeds: building realistic, reproducible evaluation environments — including virtual marketplaces and cross-organization workflows — to evaluate, compare, and accelerate progress in multi-agent research¹.

The second is the science of agent networks, aimed at understanding the safety-relevant properties of groups of interacting agents, including the mechanisms by which collective capabilities and failures emerge¹.

The third is strengthening agent infrastructure: stress-testing the protocols for agent identity, reputation, and commitment to secure interactions across platforms¹.

The fourth is oversight and control, developing methods to monitor deployed agent populations and mitigate collective harms at scale¹.

The Flip Side of Mass Agent Deployment

The announcement comes as the industry moves toward deploying agents at scale. Google DeepMind itself made agent-based tools a centerpiece of Google I/O in May 2026². In the developer tools industry, product pivots and acquisitions premised on long-running agents have been frequent, and configurations where multiple agents work in parallel in the same environment are becoming commonplace.

A company that sells agents funding research into the risks of their interactions can be read as an acknowledgment that safety research has not kept pace with deployment. James Fox, who leads the Trustworthy AI program at Schmidt Sciences, also spoke to MIT Technology Review about the need for this research².

Applications for the research funding are being accepted through the Schmidt Sciences portal, with a deadline of August 8, 2026¹. Program details and application requirements are available in DeepMind’s announcement¹ and on the application page.