Google DeepMind, alongside Schmidt Sciences, the Cooperative AI Foundation, the Advanced Research and Invention Agency (ARIA), and with support from Google.org, has announced a technical research funding call of up to $10 million targeting the safety challenges posed by multi-agent AI systems. Proposals are due August 8, 2026, with awards expected to be announced in autumn of that year.

The Core Problem: Emergent Behavior at Scale

The initiative addresses a gap that conventional AI safety evaluations have largely ignored: what happens when millions of independent AI agents, built by different organizations, interact across shared digital environments. Current safety assessments typically analyze models in isolation. As agents begin communicating, negotiating, and transacting with one another at scale, collective behaviors can emerge that no single model evaluation would predict or catch.

The organizing coalition frames this as an urgent, underexplored risk. Unanticipated population-level behaviors could trigger cascading economic disruptions or open new classes of security vulnerabilities. DeepMind points to its own prior work on “AI Agent Traps” as an example of adversarial vulnerabilities agents face in these environments, and notes that a 2025 internal framework began mapping the interaction space, but argues that complexity is now outpacing existing models.

Four Priority Research Areas

  • Sandboxes and testbeds: Developing realistic, reproducible environments for evaluating multi-agent safety, including virtual marketplaces, simulated ecosystems, and multi-organization workflow simulations.
  • Science of agent networks: Investigating how collective capabilities emerge and scale, how agent populations fail or become volatile, and how to detect dangerous population-level properties before they propagate.
  • Strengthening agent infrastructure: Stress-testing identity, reputation, and commitment protocols that underpin secure cross-platform agent interactions.
  • Oversight and control: Building methods to monitor deployed agent populations and intervene against collective harms at scale.

Organizational Alignment

The funding call connects several parallel institutional efforts. Schmidt Sciences contributes through its Science of Trustworthy AI and AI Agents programs, which focus on foundational risk research for frontier systems. ARIA’s Scaling Trust programme is oriented toward enabling secure cyber-physical multi-agent coordination. The coalition’s stated rationale is that no single laboratory can adequately address multi-agent safety, making a distributed, global research community essential for producing standards that are transparent and broadly applicable.

Researchers interested in submitting proposals can find technical requirements and application details through the organizers’ application portal.