Millions of AI agents built by different organizations will soon communicate, negotiate, and transact with each other—and we currently lack the tools to predict the collective behaviors that could emerge. Google DeepMind, together with Schmidt Sciences, the Cooperative AI Foundation, ARIA, and supported by Google.org, is throwing up to $10 million at that gap.
The Emergent Risk Gap
Most safety evaluations today test a single model in isolation. That’s fine for a chatbot, but useless when autonomous agents from competing firms start trading on a shared marketplace or negotiating access to compute. As DeepMind’s own 2025 framework and their recent work on AI Agent Traps showed, interacting agents produce emergent behaviors that are hard to anticipate. The funding call explicitly targets the “invisible” safety risks that appear only when independent systems interact across networks.
This is not theoretical. Large-scale multi-agent systems can trigger unpredictable economic flurries or create novel security vulnerabilities. We don’t have benchmarks for population-level volatility, nor tools to detect dangerous phase transitions in agent networks. The core objective of this initiative is to build exactly that capability.
Four Priority Areas for a New Safety Science
The call for proposals carves out four concrete research tracks. First, sandboxes and testbeds—realistic, reproducible environments like virtual marketplaces and simulated ecosystems where multi-agent interactions can be evaluated and compared. Second, the science of agent networks—understanding how collective capabilities emerge, how networks fail or become volatile, and how to detect dangerous population-level properties. Third, strengthening agent infrastructure—stress-testing protocols for identity, reputation, and commitment that secure cross-platform agent interactions. Fourth, oversight and control—methods to monitor deployed agent populations and mitigate collective harms at scale.
Each area corresponds to a yawning hole in current safety research. No single lab can fill them alone. That’s why the consortium is funding external academics and independent researchers globally, with awards announced in Autumn 2026.
Why Now, and What Changes
The announcement explicitly states that the complexity of multi-agent interactions is outpacing existing safety models. DeepMind’s previous work lays a foundation, but the rate of deployment—agents writing code, agents bidding on ads, agents coordinating supply chains—demands a faster, broader research response. ARIA’s Scaling Trust programme and Schmidt Sciences’ Science of Trustworthy AI programs provide institutional heft.
Applications close August 8, 2026. That deadline sets the clock for researchers to propose work that could shape how the next generation of autonomous systems is controlled at scale.
Source: Investing in multi-agent AI safety research
Domain: deepmind.google
Comments load interactively on the live page.