Google DeepMind and partners put $10M behind multi-agent AI safety research

The funding call is open to researchers worldwide and focuses on the risks that may emerge when large populations of AI agents interact across shared digital systems.

Google DeepMind and partners have opened a $10 million funding call for research into the safety of interacting AI agent systems

Google DeepMind, Schmidt Sciences, the Cooperative AI Foundation, the Advanced Research and Invention Agency, and Google.org have opened a funding call worth up to $10 million for technical research into multi-agent AI safety.

The call is aimed at researchers studying what happens when large numbers of AI agents, built and deployed by different organizations, interact across digital environments. Proposals are due by August 8, 2026, with selected awardees expected to be notified in fall 2026.

The partners are seeking work that can help researchers and developers understand, evaluate, and control risks that may emerge from interacting AI agent populations, rather than from individual models operating in isolation.

The funding is open to researchers worldwide and covers projects lasting one to two years. Tier 1 grants are available for projects up to $300,000, while Tier 2 funding ranges from $300,000 to $1 million.

The next step for applicants is to submit proposals through the application portal before the August deadline. Informational webinars are scheduled for June 30 and July 23.

A shift from single models to agent networks

Most current AI safety evaluations focus on individual models. The new call focuses instead on multi-principal, multi-agent systems, where AI agents from different developers or organizations may communicate, negotiate, transact, and coordinate across shared infrastructure.

Owen Larter, senior director and head of policy and public affairs at Google DeepMind, wrote in a LinkedIn post: "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."

The call follows recent research from Google DeepMind on distributional AGI safety, which argues that advanced AI capabilities may emerge through coordinated networks of specialized systems, rather than a single monolithic artificial general intelligence system.

The funding partners also cite work from ARIA’s Scaling Trust program and the Cooperative AI Foundation’s report on multi-agent risks from advanced AI. Those strands of work point to possible failure modes including collusion, conflict, destabilizing dynamics, emergent agency, and security vulnerabilities specific to agent populations.

Four areas of research

The call is organized around four research clusters: sandboxes and testbeds, the science of agent networks, agent infrastructure, and multi-agent oversight and control.

The first cluster focuses on realistic and reproducible environments where researchers can study frontier-model agents interacting over time. The call identifies virtual marketplaces, simulated ecosystems, and multi-organization workflows as examples of testbeds that could support comparative evaluation.

The second cluster asks researchers to examine the safety-relevant properties of agent networks, including how collective capabilities emerge, how networks become volatile or fail, and how dangerous population-level behaviors can be detected.

The third cluster focuses on the technical infrastructure that could support safer interaction between agents. Areas include identity, authentication, reputation, provenance, verifiable attributes, commitments, and accountability.

The fourth cluster covers oversight and control. The call asks for methods to monitor deployed agent populations, detect undesirable coordination, attribute failures to agents or delegation chains, and intervene when collective harms start to emerge.

What is out of scope

The funders are not seeking proposals focused only on single-agent alignment, interpretability, robustness, jailbreak defenses, or prompt injection protections.

The call also excludes work primarily aimed at increasing AI agent capabilities, AI systems that help humans cooperate with each other, non-technical policy or philosophy projects, toy systems that do not engage with frontier-model agents, and commercial product development.

Projects will be assessed on research fit, scientific quality, potential impact, feasibility, team expertise, cost, and philanthropic fit. The call says proposals should address technical work that markets are unlikely to solve on their own.

The application deadline is August 8, 2026, at 11:59 pm Anywhere on Earth. Award decisions are expected in fall 2026.