---
title: "Google DeepMind Drops $10M to Study AI Agents Talking to Each Other"
subtitle: "Google DeepMind and four partners launch a $10M fund to study what happens when millions of AI agents interact — a safety problem no one has solved yet."
url: "https://alphasignal.ai/news/google-deepmind-drops-10m-to-study-ai-agents-talking-to-each-other"
author: "AlphaSignal Newsroom"
author_url: "https://alphasignal.ai/editorial-team"
company: "Google DeepMind"
company_url: "https://alphasignal.ai/company/googledeepmind"
published: "2026-06-11T11:19:58.000Z"
updated: "2026-06-11T12:16:17.543Z"
type: "News"
topics:
  - "Agents"
  - "Business"
subtopics:
  - "Multi Agent"
  - "Agent Frameworks"
upvotes: 463
read_time: "6 min"
source_url: "https://deepmind.google/blog/investing-in-multi-agent-ai-safety-research/?utm_source=x&utm_medium=social&utm_campaign=&utm_content="
comments: 0
---

# Google DeepMind Drops $10M to Study AI Agents Talking to Each Other

> Google DeepMind and four partners launch a $10M fund to study what happens when millions of AI agents interact — a safety problem no one has solved yet.

[Google DeepMind](https://deepmind.google/blog/investing-in-multi-agent-ai-safety-research/) has teamed up with Schmidt Sciences, the Cooperative AI Foundation, the UK government's Advanced Research and Invention Agency (ARIA), and Google.org to launch a $10 million research fund targeting one of AI's least-studied risks: what happens when millions of autonomous agents start talking to each other.

## A Problem That Barely Exists Yet , But Will Soon

For years, AI safety research has focused on individual models. Can this model be jailbroken? Does it follow instructions? Is it aligned with human values? The emergence of advanced multi-agent systems presents risks that have been systematically underappreciated and understudied, in part because the deployment of such systems is currently rare, or constrained to highly controlled settings.

Google DeepMind, which made agent-based tools a centerpiece of Google I/O last month, has teamed up with several other organizations to announce a $10 million funding pot for researchers to study the behavior of multi-agent systems and come up with ways to prevent unsafe scenarios. Rohin Shah, who directs DeepMind's AGI safety and alignment research, says the window to get ahead of this is narrow. He estimates we have a few more months before agents are deployed throughout the economy in numbers that make potential risks a real concern.

## Why Multi-Agent Safety Is Its Own Beast

The core insight driving this fund is that safety problems in multi-agent systems are not just harder versions of single-agent problems , they are categorically different. These risks are distinct from those posed by single agents or less advanced technologies, and will not necessarily be addressed by efforts to mitigate the latter.

Interactions between agents may introduce new failure modes or exacerbate existing vulnerabilities. Unlike single-agent architectures, where failures tend to remain localized, multi-agent systems exhibit complex interdependencies that can propagate or even amplify existing vulnerabilities. Think of it like traffic: you can make every individual car perfectly safe, and still end up with gridlock, pile-ups, or road rage at scale.

The specific failure modes researchers are worried about include:

- **Prompt injection at scale:** As agent systems scale, network effects can amplify vulnerabilities , cascading jailbreaks or prompt injections proliferating across agent boundaries. One experiment showed a single adversarial input compromising up to a million LLM agents through cascading interactions.
- **Collusion:** Seemingly benign agents might establish secret collusion channels through steganographic communication, engage in coordinated attacks that appear innocuous when viewed individually, or exploit information asymmetries to covertly manipulate shared environments such as markets or social media.
- **Emergent instability:** These risks manifest through complex interactions and the emergent behaviors that arise from these systems' unprecedented complexity, including miscoordination, conflict, and collusion.

Shah's framing is deliberately grounded. The near-term risks are mostly supercharged versions of problems that already exist online: scams, cyberattacks, and prompt injections. But the concern is that agent-to-agent interaction creates feedback loops that make these problems orders of magnitude harder to contain. This reframes AI safety to reflect a more networked future , where risks don't just come from rogue individual systems, but from interaction failures among many. Overlooking multi-agent dynamics would be like regulating individual cars without considering traffic.

## Who's Behind the Fund

Joining Google DeepMind are Schmidt Sciences, a philanthropic foundation set up by Eric and Wendy Schmidt; ARIA, the UK government's moonshot agency; the Cooperative AI Foundation, a UK-based nonprofit research outfit; and Google's charitable arm, Google.org. The coalition is notable for spanning industry, philanthropy, academia, and government , a deliberate signal that no single lab can own the safety standards the rest of the world has to trust.

The Cooperative AI Foundation is not a new player here. Their technical report offers one of the first comprehensive attempts to map the risks that emerge when multiple advanced AI agents interact, adapt, and evolve together. DeepMind itself [published foundational framework research](https://arxiv.org/abs/2512.16856) on multi-agent interactions, and their recent work on AI Agent Traps explores vulnerabilities agents face in adversarial environments.

## Four Research Priorities

The fund is structured as an open call for proposals from academic and independent researchers worldwide. Four priority areas are defined:

1. **Sandboxes and testbeds** , Building realistic, reproducible environments to evaluate multi-agent safety, including virtual marketplaces, simulated ecosystems, and multi-organization workflows.
2. **The science of agent networks** , Understanding how collective capabilities emerge and scale, how networks fail or become volatile, and how to detect dangerous population-level properties before they manifest.
3. **Strengthening agent infrastructure** , Stress-testing the protocols for identity, reputation, and commitment that underpin cross-platform agent interactions.
4. **Oversight and control** , Developing methods to monitor deployed agent populations and mitigate collective harms at scale.

The deadline to apply is August 8, 2026, with awardees expected to be announced in Autumn 2026. Applications go through Schmidt Sciences' [application portal](https://schmidtsciences.smapply.io/prog/scaling_ai_safety_for_a_multi_agent_world).

## Industry Isn't Waiting

DeepMind is not alone in sounding the alarm. Anthropic published a guide titled *Zero Trust for AI Agents* on the Claude blog, focusing on the secure deployment of autonomous AI agents in corporate environments and outlining key risks associated with agent systems. According to Anthropic, advanced models have reduced the time between the discovery of a vulnerability and its exploitation from months to hours, and the company suggests considering not only AI-accelerated attacks on infrastructure but also the risks posed by the agents themselves, which can interpret goals, select tools, and perform multi-step actions without constant human involvement.

The market backdrop makes the urgency clear. The global AI agents market is projected at $10.7–10.9 billion in 2026, growing to $47–53 billion by 2030 at roughly 45–50% CAGR. Yet 88% of organizations deploying AI agents have experienced confirmed or suspected security incidents, and only 14.4% of those agents went to production with full security and IT approval.

## The Real Bet Here

$10 million is not a large number by AI lab standards. Shah acknowledges it is dwarfed by DeepMind's own internal research budgets. The goal is not to fund a solution , it is to fund a field. The science of trustworthy AI is nascent. Today, many demonstrations of unsafe behavior are isolated, and mitigations do not generalize reliably across model families, deployment contexts, or capability regimes. The aim is to support basic technical research that improves our ability to understand, predict, and control risks from frontier AI systems.

The deeper question is whether academia can move fast enough. The technical community faces an urgent coordination problem of its own: developing robust frameworks for multi-agent AI safety before widespread deployment makes after-the-fact intervention impractical. What this fund is really buying is time , and the hope that a diverse, independent research community can see around corners that industry labs, racing to ship agents, might miss.

## Sources

- [X post by @googledeepmind](https://x.com/googledeepmind/status/2065031279213441309)
- [Investing in multi-agent AI safety research — Google DeepMind](https://deepmind.google/blog/investing-in-multi-agent-ai-safety-research/)
- [Google DeepMind is worried about what happens when millions of agents start to interact — MIT Technology Review](https://www.technologyreview.com/2026/06/11/1138794/google-deepmind-is-worried-about-what-happens-when-millions-of-agents-start-to-interact/)
- [Multi-Agent Risks from Advanced AI — Cooperative AI Foundation](https://www.cooperativeai.com/post/new-report-multi-agent-risks-from-advanced-ai)
- [Science of Trustworthy AI — Schmidt Sciences](https://www.schmidtsciences.org/trustworthy-ai/)
- [AI agent safety in 2026: the complete guide — Responsible AI Labs](https://responsibleailabs.ai/knowledge-hub/articles/ai-agent-safety-2026)