Google DeepMind, together with Schmidt Sciences, the Cooperative AI Foundation, ARIA, and Google.org, announced a $10 million funding call for research into the safety of large-scale multi-agent AI systems.
The initiative targets a blind spot in current AI safety research: most evaluations analyze models in isolation, but the real-world deployment of millions of interacting AI agents — built by different organizations and communicating across digital environments — could produce unpredictable collective behaviors. DeepMind’s own 2025 research established a framework for understanding these interactions, while recent work on “AI Agent Traps” explored vulnerabilities agents face in adversarial environments.
The funding call focuses on four priority areas: building realistic sandboxes and testbeds for multi-agent evaluation; understanding the safety-relevant properties of interacting agent populations; stress-testing protocols for identity, reputation, and commitment in cross-platform agent interactions; and developing methods to monitor deployed agent populations and mitigate collective harms.
“We are at a critical juncture where the complexity of multi-agent interactions is outpacing existing safety models,” DeepMind said in its announcement. The deadline for researcher applications is August 8, 2026, with awardees expected in autumn 2026.
The initiative advances the missions of Schmidt Sciences’ Trustworthy AI program and ARIA’s Scaling Trust program, reflecting growing consensus that multi-agent safety research needs urgent, coordinated investment before agent ecosystems reach critical mass.
