DeepMind and a handful of academic partners have opened a $10 million funding call for multi-agent AI safety projects.
The program will award up to $500 000 per team for work that studies how multiple advanced agents might coordinate, compete, or unintentionally cause damage when deployed together. Applicants must outline concrete safety metrics and deliver open-source tools or datasets. The call runs through early 2027, with reviewers drawn from both academia and industry.
The push reflects growing concern that today's AI safety work, largely focused on single-agent models, overlooks the complexities of ecosystems where dozens of agents interact. If unchecked, such dynamics could amplify risks in areas like autonomous traffic, financial trading, or large language model collaborations. By funding early‑stage work, DeepMind hopes to seed standards before the problems become entrenched.
In short, the grant is a preemptive bet that shaping safety research now will keep multi-agent systems from becoming the next blind spot in AI governance.