DeepMind puts $10M behind multi-agent AI safety research

Google DeepMind announced a $10 million research funding call focused on the safety of multi-agent AI systems, run jointly with Schmidt Sciences, the Cooperative AI Foundation, ARIA, and Google.org. The framing is that as millions of agents from different organizations start to interact, negotiate, and transact with one another, the unsolved problems shift from "is this one model aligned" to "how does the population behave."

The call lists four priorities. The first is sandboxes and testbeds, including realistic environments like simulated marketplaces. The second is agent network science: how collective capabilities emerge in agent populations, how networks become volatile, and how to detect dangerous population-level properties before they cause harm. The third is infrastructure, meaning stress-testing protocols for identity, reputation, and commitment that secure cross-platform agent interactions. The fourth is oversight: monitoring deployed agent populations and mitigating collective harms at scale.

DeepMind frames it as a problem no single lab can solve, which is unusually direct for a Google announcement. Applications close August 8, 2026, with awards expected in autumn 2026. The funded work is meant to produce both research output and tooling that other labs can build on.

Why it matters

If you are building any product where agents talk to each other across organizational boundaries, this is the research bet to track. The funded priorities double as a fairly explicit map of where DeepMind thinks the unsolved problems sit, which is useful when deciding what to build in-house and what to wait on.