Investing in multi-agent AI safety research

Google DeepMind isn't just worried about one chatbot going rogue anymore. Alongside Schmidt Sciences, the Cooperative AI Foundation, Britain's ARIA agency, and Google.org, the lab announced on June 11 a technical research fund worth up to $10 million aimed at a much messier problem: what happens when millions of AI agents, built by totally different companies, start talking, trading and negotiating with each other across the open internet.

The pitch here is that a decade of AI safety work has mostly focused on making single models behave. But agents don't stay single for long. DeepMind's own researchers point out that when autonomous systems start interacting at scale, you get emergent behavior — collective dynamics that nobody designed and nobody can easily predict. Think sudden bursts of automated economic activity, or new attack surfaces that only appear once thousands of agents are transacting simultaneously. Nobody currently has good tools to spot these shifts before they happen, let alone stop them.

The money is meant to plug that gap, and it comes with four specific asks. Researchers can build sandboxes and testbeds — simulated marketplaces and multi-organization workflows — to stress-test agent behavior safely. Others can dig into the actual science of agent networks: how capabilities scale across populations, how networks destabilize, how to catch dangerous group-level patterns early. There's a call to harden the plumbing of agent-to-agent interaction itself, meaning identity verification, reputation systems and commitment protocols that would let agents trust each other across platforms. And finally, there's oversight — building monitoring systems that can watch deployed agent populations and catch collective harm before it spreads.

This builds on work DeepMind published in 2025 on multi-agent interaction frameworks, plus more recent research on what it calls