Google DeepMind is worried about what happens when millions of agents start to interact

Google DeepMind, together with Schmidt Sciences, ARIA, and others, is putting $10 million into academic research on multi-agent safety. Rohin Shah, who leads AGI safety at DeepMind, says a dedicated research field for multi-agent safety doesn't exist yet and they want to help build one. The fear isn't a single rogue agent—it's that millions of agents interacting online could supercharge existing problems like scams, prompt injections, and cyberattacks. Shah estimates we are still months away from mass deployment and wants sandbox simulations ready before that. The post does not disclose application criteria or selection timelines.

Why it matters: DeepMind plus external funders are putting real money into multi-agent safety as a field, with Rohin Shah on the record. Not a product update, but the topic is forward-looking and directly relevant to teams building agents. Score isn't higher because it's a funding announcemen...