Investing in multi-agent AI safety research

Overview

Google DeepMind and four partners opened a research funding call worth up to $10 million to study how AI agents behave when many of them interact at scale. The money targets safety risks that surface only when independent agents built by different organizations work together, a gap most current safety testing ignores. Academic and independent researchers worldwide can apply by August 8, 2026, with award decisions expected in autumn 2026.

Key Takeaways

Google DeepMind, Schmidt Sciences, the Cooperative AI Foundation, the UK government's Advanced Research and Invention Agency (ARIA), and Google.org are jointly funding the program with a combined commitment of up to $10 million.
The grants fund technical research into safety problems that appear only when large numbers of AI agents from separate systems interact at once, rather than single models tested in isolation.
Applications are open to academic and independent researchers anywhere in the world, submitted through the Schmidt Sciences SMApply system.
The application deadline is August 8, 2026, and the partners plan to announce which projects receive funding in autumn 2026.
Funding is split across four research tracks: test environments, the science of agent networks, agent infrastructure, and oversight and control.
Rohin Shah, who leads Google DeepMind's AGI safety and alignment research, framed agents acting on instructions from other agents without human oversight as a new class of risk.

Stats & Key Facts

#Up to $10 million in total research grants committed by the five partners.
#5 organizations are backing the funding pool: Google DeepMind, Schmidt Sciences, the Cooperative AI Foundation, ARIA, and Google.org.
#4 priority research areas define how the money is allocated.
#August 8, 2026 is the single application deadline for all proposals.
#Award decisions are expected in autumn 2026, roughly a few months after the deadline.
#The program is open to researchers across every country, not limited to one region.

A $10 Million Pool Backed by Five Organizations

The funding brings together a research lab, two science foundations, a government agency, and a philanthropic arm.

›Google DeepMind leads the call alongside Schmidt Sciences and the Cooperative AI Foundation.
›The UK government's Advanced Research and Invention Agency (ARIA) and Google.org round out the five partners.
›The combined commitment totals up to $10 million in grants for technical research.
›Proposals are submitted through the Schmidt Sciences SMApply application portal.

Why Many Agents Interacting Creates New Safety Risks

Most AI safety testing today checks one model on its own. The new program argues that approach misses a growing problem: real deployments increasingly involve many autonomous agents, often built by different companies, working alongside each other on the same tasks.

When independent agents interact, they produce group behaviors that are hard to predict from studying any single agent. Failures can spread from one agent to another across a network, and dangerous patterns can emerge at the population level that no one designed or expected.

Rohin Shah, who directs AGI safety and alignment research at Google DeepMind, described the core concern as agents carrying out tasks without human oversight and following instructions handed to them by other agents. He framed this combination as a fresh category of risk that existing safety work does not cover.

Four Research Tracks the Money Funds

The grants are organized around four specific problem areas.

›Sandboxes and testbeds: reproducible environments such as virtual marketplaces, simulated ecosystems, and multi-organization workflows for evaluating agent behavior safely.
›Science of agent networks: how collective capabilities emerge in agent populations, how those networks fail, and how to detect harmful population-level properties.
›Strengthening agent infrastructure: stress-testing protocols for identity, reputation, and commitment so agents from different platforms interact securely.
›Oversight and control: scalable methods to monitor deployed agent populations and reduce harm when something goes wrong.

Who Can Apply and the Key Dates

The call is open broadly, with a firm timeline.

›Eligible applicants: academic researchers and independent researchers worldwide.
›Application deadline: August 8, 2026.
›Decision announcement: autumn 2026.
›Submission method: the Schmidt Sciences SMApply system.

How This Connects to Existing Safety Programs

The call extends work the partners already fund elsewhere.

The initiative feeds into two Schmidt Sciences efforts, its Science of Trustworthy AI program and its AI Agents program, both aimed at making AI systems more reliable and accountable. It also advances ARIA's Scaling Trust programme, a UK government push to develop tools that keep advanced AI dependable as it grows.

By tying the new grants to these existing programs, the partners signal a coordinated push rather than a one-off grant. The structure pools money and direction from a research lab, foundations, and a government agency toward the same multi-agent safety goal.

What This Means for Businesses Adopting AI Agents

Companies are moving quickly to deploy autonomous AI agents that handle tasks with little human supervision. As adoption spreads, agents from different vendors will start coordinating on shared workflows, from procurement to customer service to logistics.

The research this funding supports aims to produce practical tools, test environments, and monitoring methods that organizations use to keep those agent interactions safe and predictable. The goal is to understand the failure modes before millions of agents operate together in production, not after problems appear in live systems.

Frequently Asked Questions

How much money is available and who is providing it?

The funding call offers up to $10 million in total. It is backed by five organizations: Google DeepMind, Schmidt Sciences, the Cooperative AI Foundation, the UK government's ARIA, and Google.org.

What kind of research does the call fund?

It funds technical research into the safety of large-scale multi-agent AI systems, meaning the risks that appear when many independent agents interact at once. Proposals fall into four areas: test environments, the science of agent networks, agent infrastructure, and oversight and control.

Who is eligible to apply?

Academic researchers and independent researchers anywhere in the world can apply. Applications are submitted through the Schmidt Sciences SMApply system.

What is the deadline and when are decisions made?

The application deadline is August 8, 2026. The partners expect to announce which projects receive funding in autumn 2026.

Why focus on multi-agent safety specifically?

Most safety testing examines one model alone, but real deployments increasingly involve many agents from different organizations working together. Those interactions create emergent group behaviors and failures that spread across networks, which single-model testing does not catch.

The $10 million call signals that leading AI labs and governments see coordinated agent behavior as the next safety frontier, funding the research before large agent populations reach broad deployment. Researchers worldwide have until August 8, 2026 to submit proposals.