Our Take
The funding targets a real gap—current safety work treats models in isolation, but deployed agents will interact in ways we cannot yet predict or measure—but the framing sidesteps the harder question: who builds the enforcement infrastructure when agents fail at scale.
Why it matters
Within two years, independent organizations will deploy AI agents that negotiate, transact, and coordinate across networks without human oversight between each interaction. Safety research that starts now shapes whether that ecosystem becomes stable or chaotic.
Do this week
Research leads: if your lab models multi-agent failure modes or builds testbed infrastructure, draft a proposal by July 15 so you can submit before the August 8 deadline.
DeepMind and partners announce $10M in multi-agent safety grants
Google DeepMind, Schmidt Sciences, the Cooperative AI Foundation, the Advanced Research and Invention Agency, and Google.org are opening a research funding call on June 11, 2026 for work on multi-agent AI safety. The call awards up to $10 million to researchers globally (per DeepMind's blog). The deadline is August 8, 2026, with winners announced in autumn 2026.
The funding targets four areas: sandboxes and testbeds for evaluating multi-agent systems; understanding how collective agent capabilities emerge and fail; securing protocols for cross-platform agent identity and reputation; and methods to monitor and mitigate harms from deployed agent populations.
DeepMind cites its own 2025 research establishing a framework for multi-agent interactions and recent work on "AI Agent Traps," which explores vulnerabilities in adversarial settings. The institute argues the complexity of multi-agent systems is outpacing existing safety models.
The real problem is emergence at scale, not safety in isolation
For a decade, AI safety work has centered on individual models. DeepMind's shift is structural: when millions of independent agents built by different organizations interact across networks—trading, negotiating, coordinating—new collective behaviors can emerge suddenly. Current safety evaluations cannot predict or measure these transitions.
The stakes are concrete. Unpredictable economic activity spikes, cascading failures across dependent systems, or new attack surfaces that no single-agent test would reveal. DeepMind frames this as a "critical juncture" where system complexity has already outpaced safety tools.
The call is also a signal about governance. No single lab can enforce safety standards once agents are deployed by competitors. Building a transparent, globally-distributed research community now means safety standards have a chance of being adopted as defaults rather than retrofitted after incidents.
Researchers: identify your gap and propose before August
The four focus areas map to real problems. If your lab builds simulation environments for agent interaction, you have a testbed proposal. If you model how agent populations fail or become volatile, that is a science-of-networks proposal. If you work on decentralized identity or reputation protocols, you have an infrastructure proposal. If you have methods to monitor and steer deployed agent behavior in real time, you have an oversight proposal.
The funding also signals that multi-agent AI is no longer theoretical. Organizations are already planning to deploy agents. Research submitted now will influence how those systems are built, not just studied after launch.