Google DeepMind opens a $10M call for multi-agent AI safety research

Google DeepMind, together with Schmidt Sciences, the Cooperative AI Foundation, the Advanced Research and Invention Agency (ARIA), and Google.org, has announced a funding call of up to $10M for external researchers working on multi-agent AI safety. This isn't a model release or a paper — it's a grant program. Proposals are due August 8, 2026, with awardees announced in autumn 2026.

The framing in the announcement is worth taking at face value: for a decade, the field's safety work has centered on making individual models more capable and safe in isolation. That's no longer the operative unit. Soon, millions of agents built by different organizations will be interacting across shared digital environments — negotiating, transacting, forming and breaking commitments — and almost none of our safety tooling was built to evaluate that.

Why single-model evaluation stops being sufficient

Most safety evaluation today still asks: does this model, on its own, behave well? That question doesn't capture what happens when hundreds of independently-built agents start interacting. DeepMind's own prior work — a 2025 framework for multi-agent interactions and their more recent research on AI Agent Traps — has already shown that populations of agents can produce behaviors that don't show up in any single agent's evaluation: emergent coordination, adversarial exploitation between agents, or system-wide instability that only appears at scale. The announcement calls these "invisible" risks precisely because they live in the interaction graph, not in any one node.

That's a real methodological gap, not a hypothetical one. If your safety evaluation only ever tests one model against a fixed set of prompts, you have no instrument for detecting what happens when that model is one of ten thousand agents transacting in a shared marketplace, each optimizing locally.

What's actually being funded

The call names four priority areas:

• Sandboxes and testbeds — reproducible environments (virtual marketplaces, simulated ecosystems, multi-organization workflows) to actually run these experiments instead of theorizing about them.
• The science of agent networks — understanding how collective capabilities emerge and scale, how networks fail or go unstable, and how to detect dangerous population-level properties before they're obvious.
• Strengthening agent infrastructure — stress-testing the protocols agents use to establish identity, reputation, and commitments across organizational boundaries.
• Oversight and control — monitoring methods for deployed agent populations and mitigating collective harms at scale, not just per-agent misbehavior.

That list maps cleanly onto the actual failure modes people building multi-agent systems run into today: no shared identity or reputation layer across vendors, no way to simulate what a marketplace of agents does before deploying it, and monitoring that's built for single-request tracing rather than population-level drift.

An informed take

Having built and run agentic systems in production, I'd flag the infrastructure track as the one most likely to matter soonest. Sandboxes and network science are necessary research, but they're also the kind of problem academia is already reasonably equipped to study. Identity, reputation, and commitment protocols across organizational boundaries are not — that's a coordination problem closer to standards-setting than to research, and right now there isn't an equivalent of TLS or OAuth for agent-to-agent trust. Every team building multi-agent pipelines today is inventing its own ad hoc version of "can I trust what this other agent just told me," usually with no formal guarantees at all.

The emergent-behavior risk is also not abstract. Anyone running a pipeline with more than two or three cooperating agents has seen small local optimizations compound into something the designer didn't intend — a retry loop that turns into runaway API spend, or two agents that each defer to the other's judgment and produce a deadlock. Those are toy-scale versions of exactly the population-level instability this call is trying to fund research on. Scaling that intuition to thousands of independently-built agents interacting across organizations is a genuinely hard, underexplored problem, and $10M spread across a global pool of academic and independent researchers is a reasonable first move — not a solution, but the right kind of bet given how early this field is.

If you do multi-agent safety research and this fits your work, the application portal and full call details are linked from the original announcement, with proposals due August 8, 2026.