Google DeepMind and partner organizations are funding $10 million（約16億円） in research to understand risks when multiple AI agents interact with each other, concerned that unsafe scenarios could become real as agent deployment scales up.

What happened: Google DeepMind, Schmidt Sciences, ARIA (the UK government's moonshot agency), the Cooperative AI foundation, and Google.org announced a $10 million（約16億円） funding pot for researchers to study multi-agent systems and develop ways to prevent unsafe scenarios.

Why it matters: As AI agents get deployed in larger numbers and begin working together, risks that are currently hypothetical—such as scams, prompt injections (malicious instructions fed to AI agents), and cyberattacks—could become real-world problems. The concern is that institutions made up of many interacting agents could accomplish things no single agent could, much like human institutions. Currently there is no established field of research for multi-agent safety.

What to watch: Researchers plan to run realistic simulations by dropping AI agents into sandboxes to study their behavior. Shah believes there are a few more months before agents are deployed throughout the economy in numbers that would make potential risks a genuine concern, and he wants to get ahead of that moment.