Google DeepMind and its partners are putting $10 million into a safety problem that gets less attention than model benchmarks: what happens when AI agents start dealing with each other at scale.
The funding call opened on June 11, 2026, and it is aimed at a narrow problem. Google DeepMind, Schmidt Sciences, the Cooperative AI Foundation, ARIA, and Google.org are asking researchers to study multi-agent AI safety, with proposals due August 8 and awards expected in fall 2026. Smaller projects can receive up to $300,000. Larger ones can receive between $300,000 and $1 million.
You should care about the exact target here. This isn't another broad AI safety pot where every concern gets folded into one vague debate about alignment. The call is about systems of AI agents interacting with other AI agents, negotiating, competing, coordinating, and sometimes acting against each other in shared digital spaces. That is a different safety problem from one chatbot giving one answer to one person.
Also read: NextEra Energy is becoming the infrastructure layer that powers the AI economy • OVHcloud is training frontier AI models and the Anthropic export ban just made its pitch a lot easier • Swiss AI startup Prem is raising $100 million as Anthropic's export ban makes on-premise infrastructure look essential
The industry has been moving fast toward agentic AI because the pitch is obvious. A model that can search, plan, click, book, write, buy, sell, and revise work on your behalf is more useful than a model that waits for a prompt and stops at text. But once you have many of these systems operating at once, the risk is no longer only whether one agent follows one instruction. The risk is what they learn to do around each other.
Frankly, this is the part of the agent story that has been too easy to wave away. A single assistant making a mistake is annoying. Thousands or millions of automated agents bidding, scraping, scheduling, bargaining, or routing tasks through the same platforms could create failures that no one agent was explicitly told to cause. You don't need science fiction for that. You only need software systems with incentives, limited oversight, and enough autonomy to act before a human checks the work.
The hard problem is interaction
The Cooperative AI Foundation's work behind the call is focused on the messy middle between individual model behavior and broad social impact. That is where a lot of real damage can sit. Agents may collude without anyone writing a collusion rule. They may learn brittle bargaining strategies. They may amplify spam, fraud, or market pressure because the local reward tells them to win and the wider system has no way to say, stop.
This is why the partner list matters. Google DeepMind brings frontier model research. Schmidt Sciences brings philanthropic money for technical work. ARIA, the United Kingdom's Advanced Research and Invention Agency, is built to fund research that doesn't always fit ordinary grant channels. The Cooperative AI Foundation has been pushing the idea that AI systems need to cooperate safely, not merely perform well in isolation.
The call's structure also tells you what kind of work they want. Grants up to $300,000 can support focused experiments, benchmarks, or theory work. Awards as high as $1 million point to larger research programs, the kind that need teams, infrastructure, and enough time to test agent behavior across repeated interactions. That is the right scale for the question. One clever paper won't settle this.
There is a practical startup angle here too, and it isn't the usual bland warning that founders should be responsible. If you're building agents for customer support, finance, procurement, sales operations, code review, or internal workflow automation, you are not just shipping a better interface. You are putting a decision-making actor into a crowded environment. Your agent may talk to another company's agent before it talks to a person. It may make choices that affect pricing, access, priority, or trust.
That changes the standard you have to meet. It isn't enough to say the model passed your internal tests when used alone. You need to know how it behaves when other agents are also optimizing, probing, refusing, copying, negotiating, or gaming the same rules. If you can't answer that, you don't yet understand the product you are selling.
The timing is current. The call opened this month, the deadline is still ahead, and the awards won't be made until fall 2026. That gives researchers a short window to turn a loose industry anxiety into testable projects. It also gives companies a warning before the agent market gets more crowded: the safety problem is moving from outputs to interactions.
Google DeepMind's $10 million won't solve that by itself. But it puts money behind the right question. Once AI agents start acting around each other at scale, the question isn't only whether they are smart. It is whether the systems they create together are ones you can live with.
