Google DeepMind and partners open $10 million funding call for multi-agent AI safety research

Google DeepMind and four partner organizations have opened a $10 million funding call for technical research into multi-agent AI safety. The initiative targets the systemic risks that emerge when large populations of independent AI agents interact across shared digital environments, rather than the risks of isolated models. Proposals are due August 8, 2026.

The shift to multi-agent systems

Current AI safety evaluations largely focus on individual models. This funding call shifts attention to multi-principal, multi-agent systems. In these environments, agents built by different organizations negotiate, transact, and coordinate, raising new safety questions for AI Agents & Automation research.

Owen Larter, senior director and head of policy and public affairs at Google DeepMind, said: "When large groups of AI agents interact, new behaviors and capabilities can emerge suddenly. From a policy and governance perspective, understanding how to manage this-and prevent unpredictable economic or security challenges-is really important."

Four targeted research clusters

The funding pool is divided into Tier 1 grants up to $300,000 and Tier 2 grants ranging from $300,000 to $1 million. Projects will last one to two years and must fall into one of four technical research clusters:

• Sandboxes and testbeds: Creating realistic, reproducible environments like virtual marketplaces or simulated ecosystems to study frontier-model interactions over time.
• Science of agent networks: Examining how collective capabilities emerge, how networks fail, and how to detect dangerous population-level behaviors.
• Agent infrastructure: Developing technical safeguards for identity, authentication, reputation, provenance, and accountability.
• Multi-agent oversight and control: Building methods to monitor deployed populations, detect undesirable coordination, and intervene when collective harms emerge.

What the funders are not looking for

Applicants must avoid proposals centered on single-agent alignment, interpretability, model resilience, or prompt injection defenses. The partners also reject work aimed primarily at increasing agent capabilities, non-technical policy projects, toy systems, or commercial product development.

Funders will assess proposals based on research fit, scientific quality, potential impact, and team expertise. They specifically seek technical work that free markets are unlikely to solve independently. Informational webinars are scheduled for June 30 and July 23 to help teams structure their submissions.

Why this matters for science and research professionals

For researchers in this field, this funding represents a direct pivot away from isolated model evaluation toward complex, systemic network behavior. It offers a rare opportunity to secure up to $1 million for foundational work in AI for Science & Research that addresses emergent, population-level risks before they scale.