DeepMind opens a $10M call for multi-agent safety, with applications closing today
Google DeepMind, Schmidt Sciences, the Cooperative AI Foundation and ARIA are funding up to $10 million of research into what happens when agents built by different organisations negotiate and transact with each other. The deadline is 8 August 2026; awards are expected in the autumn.
The framing is the notable part. Almost all deployed safety work targets a single model responding to a single user. This call targets the interaction layer — millions of agents from different vendors, with different objectives and no shared operator, communicating and transacting in the same environments.
That is a different failure surface. A system can be individually well-behaved and still produce collusion, cascading failure or exploitable equilibria in a population. None of the standard single-model evaluations look for it, and the funders are explicit that the opportunity is to build the safety properties in while the ecosystem is still forming rather than retrofit them later.
It is also a quietly commercial problem. The security market is already pricing agent-to-agent risk ahead of the research that would tell anyone how to measure it.
Google DeepMind — Investing in multi-agent AI safety research → · arXiv — International AI Safety Report 2026 →