Google DeepMind funds $10M research into risks as millions of AI agents prepare to interact

Google DeepMind Warns of New Risks as AI Agents Scale

Google DeepMind is sounding the alarm about what happens when millions of independent AI agents interact across digital environments. The company has partnered with Schmidt Sciences, the Cooperative AI Foundation, ARIA, and Google.org to announce a $10 million multi-agent safety research funding call for researchers worldwide 1

1

2

2

. According to Rohin Shah, who directs Google DeepMind's AGI safety and alignment research, the mass deployment of AI agents that can carry out tasks without human oversight and follow instructions from other agents creates an entirely new class of risk 1

1

.

The initiative comes as Google DeepMind made agent-based tools a centerpiece of Google I/O last month, signaling the technology's imminent arrival 1

1

. Shah estimates we have just a few more months before agents are deployed throughout the economy in numbers that make potential risks a real concern. The goal is to get ahead of that moment by kickstarting AI safety research outside of tech companies.

Understanding the Risks of Multi-Agent AI Systems

The risks of multi-agent AI systems that Shah and James Fox, who leads the Science of Trustworthy AI program at Schmidt Sciences, have in mind are largely supercharged versions of existing internet threats. These include scams, prompt injections where an AI agent is fed malicious instructions and turned into self-guiding malware, and other forms of cyberattacks . "We've got this digital commons that is integral to how society works and you really want to ensure that this doesn't descend into just absolute anarchy," Fox explained 1

1

.

Beyond immediate threats, there's concern about unpredictable economic activity and security challenges that could emerge when large-scale multi-agent AI systems begin interacting. Shah noted that just as human institutions can accomplish things no individual human can, AI agents working together could hit a tipping point where imagined scenarios become real 1

1

. When asked about worst-case scenarios like widespread economic collapse, Shah didn't dismiss the possibility, though he suggested it wouldn't happen by year's end.

Building a Field for Multi-Agent Safety

"The main issue is that there just isn't really a field of research for multi-agent safety yet," Shah stated. "And we would like there to be" 1

1

. While the $10 million funding is dwarfed by Google DeepMind's own research budgets, the aim is to leverage academia's strength in looking far into the future and conducting work that isn't top of mind at industry labs 1

1

.

The funding call focuses on four priority areas: building sandboxes and testbeds for realistic evaluation, understanding the science of agent networks and how collective capabilities emerge, strengthening agent infrastructure including protocols for identity and reputation, and developing oversight methods to monitor deployed agent populations 2

2

. Both Shah and Fox emphasize that the only way to understand what happens when independent AI agents interact is through realistic simulations, as you cannot predict behavior by studying single agents or small groups in isolation 1

1

.

Why This Matters Now

The complexity of agent network dynamics stems from having huge numbers of interactions happening simultaneously, and AI agents underpinned by large language models cannot be assumed to always act rationally 1

1

. Rafael Angel, cofounder and CTO of cybersecurity firm Akeyless, noted that every past approach to security assumed machines were software written by humans, doing fixed things on fixed paths. "An agent breaks all of those assumptions: it reasons, it improvises, and it can be hijacked by a single sentence buried in a document it was asked to read," Angel explained 1

1

.

Google DeepMind's recent work established frameworks for understanding these interactions and explored vulnerabilities agents face in adversarial environments 2

2

. The company acknowledges we are at a critical juncture where the complexity of multi-agent interactions is outpacing existing safety models. Some researchers at Google DeepMind have even argued that artificial general intelligence could emerge not from a single super-smart model but from a kind of agent hivemind, where the capabilities of the whole exceed the sum of its parts 1

1

.

The deadline to apply for this multi-agent safety research funding is August 8, 2026, with awardees expected to be announced in Autumn 2026 2

2

. By supporting a global network of independent researchers, the initiative aims to ensure safety standards for trustworthy AI systems are transparent and robust for everyone as we enter this new era of AI deployment.