Google DeepMind and unnamed partners are putting $10 million on the table for researchers willing to study what happens when AI agents don't behave.
The funding call invites proposals focused on multi-agent AI safety, the study of risk that emerges not from a single model but from systems where multiple agents interact, negotiate, or compete. DeepMind is co-sponsoring with unspecified partners, framing the initiative as an external grant program rather than in-house lab work. The $10M pool is meant to seed independent research on coordination failures, emergent deception, and other hazards that single-agent safety frameworks were never built to catch.
Multi-agent deployments are already shipping. Several major AI labs, including Google, are building agentic pipelines where models hand tasks to other models, systems that can fail in ways no individual component would predict. Most safety research still treats AI as one system talking to one human; a world of agents talking to agents demands a different set of guarantees.
Ten million dollars is a modest number given the pace of deployment, but it is telling that DeepMind is funding outside researchers to scrutinize a category of technology it is also actively building.