Google DeepMind Launches $10M Multi-Agent AI Safety Fund

On June 11, 2026, Google DeepMind announced a research funding initiative of up to $10 million aimed squarely at a problem the AI industry has largely talked around: what happens when millions of autonomous agents — built by different organizations, running on different platforms — start negotiating, transacting and communicating with one another at scale. The program, formally called “Scaling AI Safety for a Multi-Agent World,” is a joint effort with Schmidt Sciences, the Cooperative AI Foundation, the Advanced Research and Invention Agency (ARIA) and Google.org.

The announcement marks a deliberate shift in how safety research is framed. For most of the past decade, AI safety work has focused on individual models — evaluating a single system’s outputs, alignment and failure modes in isolation. That approach increasingly misses the point as agent-to-agent interactions become a standard part of how AI is deployed in the real world. The concern isn’t just that one agent might behave badly; it’s that populations of agents can produce unpredictable collective behaviors that no single actor designed or anticipated.

The funding call is structured around four priority research areas: building realistic sandboxes and testbeds to evaluate multi-agent systems; studying how collective capabilities emerge and how agent networks fail; stress-testing the identity and reputation protocols that underpin cross-platform agent interactions; and developing tools to monitor and intervene in deployed agent populations. Project grants will run one to two years, with Tier 1 awards up to $300,000 and Tier 2 awards ranging from $300,000 to $1 million. Applications are open to researchers worldwide, with proposals due August 8, 2026, and awardees expected to be notified in fall 2026.

The initiative builds on prior work from DeepMind’s own research pipeline. Its 2025 framework laid early groundwork for understanding multi-agent interactions, and more recent research on “AI Agent Traps” documented specific vulnerabilities agents face in adversarial settings. The new call is an acknowledgment that internal research alone isn’t sufficient — and that the complexity of the problem is now outpacing existing safety models.

How This Fits Into the Broader Safety Landscape

The announcement doesn’t exist in a vacuum. Anthropic recently introduced a zero-trust deployment framework for AI agents, and its AI Safety Fellows Program is currently accepting applicants for cohorts beginning in May and July 2026, offering four-month residencies with a $3,850 weekly stipend and roughly $15,000 per month in compute. The AI Safety Fund, backed by the Frontier Model Forum with support from Anthropic, Google, Microsoft, OpenAI and others, including Schmidt Sciences, also funds independent safety research at a broader level.

What distinguishes the new DeepMind-led call is its specific focus on multi-principal, multi-agent dynamics — the failure modes that arise not from any single agent’s misbehavior but from the system-level interactions between many of them. The Cooperative AI Foundation’s published research on multi-agent risks identifies a distinct cluster of threats in this space, including collusion between agents, destabilizing emergent behaviors, and novel security vulnerabilities that have no analog in single-agent systems. ARIA’s Scaling Trust program, one of the co-funders, is explicitly oriented around unlocking safe forms of cyber-physical multi-agent coordination — suggesting this isn’t purely theoretical.

The urgency is underscored by recent events. On June 10 — one day before the announcement — Microsoft disabled 73 GitHub repositories after a malware attack that specifically targeted automated coding routines, a reminder that agent-to-agent infrastructure is already being exploited in the wild.

There’s also a structural distinction worth noting: unlike Anthropic’s Fellows Program, which requires applicants to have work authorization and reside in the United States, the UK or Canada, the DeepMind call imposes no geographic restrictions. That matters for researchers at institutions outside those three countries who are working on equally relevant problems.

Why This Matters If You’re a Student or Early-Career Researcher

For graduate students and postdocs, this kind of open, external grant is relatively rare in the AI safety space. Most high-profile safety roles are internal positions at major labs, which means the research agenda is set by the employer and publication timelines are subject to company priorities. This call is explicitly non-commercial — the stated intent is to fund work that markets are unlikely to produce on their own — which aligns well with academic career incentives around publication and independent credibility.

The four research areas are also broader than they might first appear. Building testbeds draws on software engineering and systems design. Studying agent network dynamics involves economics, complex systems theory and machine learning. Infrastructure stress-testing overlaps with cybersecurity. Oversight tooling requires both ML expertise and policy thinking. A CS graduate student with a systems or security background may be just as competitive as someone working in formal alignment theory.

Perhaps most importantly, multi-agent safety is a field without settled frameworks. Researchers who enter now have a genuine opportunity to define the vocabulary and evaluation standards that the rest of the field will eventually adopt — a career position that gets harder to achieve as a discipline matures.

Researchers interested in applying can review the full proposal requirements and submit their application here. The deadline is August 8, 2026.

Source: Google DeepMind