Google DeepMind launches $10M initiative to study what happens when millions of AI agents interact
For most of the history of AI safety research, the unit of analysis has been the single agent: Is this model truthful? Is it aligned with human values? Can it be jailbroken? But a growing number of researchers argue that the real risks will not come from individual AI agents acting alone — they will come from what happens when millions of them start interacting.
On June 11, Google DeepMind announced a new research initiative to address precisely that question. In partnership with Schmidt Sciences, the Cooperative AI Foundation, the UK government’s moonshot agency ARIA, and Google.org, DeepMind has opened a funding call of up to $10 million for research into the dynamics of multi-agent AI systems.
“New capabilities can arise unpredictably when many agents interact — much like how human institutions can do things that no individual person can,” said Rohin Shah, who directs DeepMind’s AGI safety and alignment research, in an interview with MIT Technology Review. “We don’t yet have the tools to predict, measure, or monitor what happens when large populations of agents operate together.”
The funding call identifies four priority research areas, each targeting a specific gap in the current understanding of multi-agent systems:
Sandboxes and testbeds. Realistic, reproducible environments where multi-agent dynamics can be studied before deployment. The goal is virtual marketplaces, simulated ecosystems, and multi-organization workflows where researchers can observe emergent behaviors under controlled conditions before they manifest in the real world.
The science of agent networks. How do collective capabilities emerge and scale? How do networks fail or become volatile? Can dangerous population-level properties be detected early? This is the most fundamental research area — analogous to the transition from individual psychology to sociology, but for artificial agents.
Strengthening agent infrastructure. Protocols for identity, reputation, and commitment across platforms. If agents are going to transact with each other, sign contracts, or delegate tasks, the underlying infrastructure needs to be stress-tested for vulnerabilities — particularly against adversarial agents that might exploit trust mechanisms.
Oversight and control. Methods to monitor deployed agent populations and mitigate collective harms at scale. This includes the technical tools for real-time observation and the governance structures for intervention when things go wrong.
Why now
The timing is not arbitrary. AI agents — systems that can act autonomously on behalf of users — are being deployed at an accelerating pace. Coding agents, customer service agents, research assistants, and automation tools are already operating in shared digital environments. The combination of rapid deployment and limited understanding of inter-agent dynamics has created what Shah describes as a “visibility gap.”
“You could have millions of agents interacting in ways that produce systemic instability without anyone noticing until it’s too late,” he said.
The article’s author, Will Douglas Heaven, notes that the concerns are not purely hypothetical. DeepMind’s own recent research on “AI Agent Traps” has explored the vulnerabilities agents face in adversarial environments. Anthropic has published zero-trust deployment guidelines for agents that assume every agent is a potential attacker. Prompt injection attacks — where an agent is fed malicious instructions hidden in otherwise benign input — have already been demonstrated to turn agents into self-guiding malware.
James Fox, who leads the Science of Trustworthy AI program at Schmidt Sciences, described the current situation more bluntly: without proactive research into multi-agent dynamics, the default outcome is “absolute anarchy in the digital commons.”
What’s next
The funding call has an August 8, 2026 deadline, with awardees to be announced in autumn. Shah’s time horizon for when multi-agent risks become a practical concern is measured in months rather than years.
“Agents are coming fast,” he told MIT Technology Review. “We need to understand the dynamics of large-scale agent interactions before they’re deployed at population scale — not after.”
Source: Heaven, W.D. “Google DeepMind is worried about what happens when millions of agents start to interact.” MIT Technology Review (June 11, 2026). Partner organizations: Google DeepMind, Schmidt Sciences, Cooperative AI Foundation, ARIA, Google.org.