Google DeepMind and Schmidt Sciences fund multi-agent AI safety research
The grant call, also backed by the Cooperative AI Foundation and ARIA, targets emergent risks from populations of interacting agents rather than any single model in isolation.
- Safety & alignment
- Minor
Google DeepMind, Schmidt Sciences, the Cooperative AI Foundation, the UK’s Advanced Research and Invention Agency (ARIA) and Google.org opened a fund of up to $10 million for academic and independent researchers studying the safety of multi-agent, multi-principal AI systems — settings where many AI agents built by different organisations interact with one another rather than a single model operating alone.
The call named four priority areas: sandboxes and testbeds that reproduce realistic multi-agent environments such as simulated marketplaces; “agent network science,” studying how collective capabilities and risks emerge from populations of interacting agents; infrastructure for agent identity, reputation and secure cross-platform interaction; and oversight and control methods for monitoring deployed agent populations at scale. The organisers argued that existing safety evaluation, which typically examines one model in isolation, cannot anticipate the emergent, system-level behaviour that arises once millions of agents from different developers begin transacting and coordinating across shared digital environments.
The initiative reflects a shift in frontier-safety research focus during 2026, from single-model alignment toward the harder problem of governing interacting populations of agents as agentic deployment — coding assistants, shopping agents, autonomous browsing tools — moved from research demonstrations into everyday products.