Google DeepMind has launched a $10 million fund aimed at studying the risks of multi-agent systems as these AI tools become more prevalent in everyday life.

Rohin Shah, who directs AGI safety and alignment research at DeepMind, explains that the main issue is the lack of a field dedicated to researching multi-agent safety. The concern is that as AI agents work together, they could potentially create dangerous scenarios similar to those seen on the internet today but amplified.

Alexander Fox from Schmidt Sciences notes that understanding these risks requires running realistic simulations, given the complexity arising from large numbers of interactions at once. The goal is to prevent potential anarchy in our digital commons by ensuring AI agents act safely and rationally.

While DeepMind’s initiative addresses some risks, other top AI firms like Anthropic are also warning about the dangers inherent in agent-based systems. Refael Angel from Akeyless highlights that traditional security approaches may not apply to these new autonomous entities, which can reason and improvise, making them harder to control.

Despite the urgency of the situation, Shah remains optimistic, believing they have a few months before significant risks arise. However, the question remains: is it enough time?