Google DeepMind funds research on risks of interaction between millions of AI agents

Google DeepMind Funds Research on Risks of Millions of AI Agents Interacting With Each Other

Google DeepMind is funding new research into the risks that emerge when millions of different AI agents interact with each other online. According to MIT Technology Review, citing Rohit Shah, who heads the AGI safety and alignment research direction at the company, the mass market launch of agents capable of performing tasks without constant human oversight and receiving instructions from other agents creates a fundamentally new class of risks that the industry has yet to fully understand.

Who is Behind the Research

Google DeepMind is Google's research division responsible for both developing advanced models in the Gemini family and conducting fundamental AI safety research. Rohit Shah oversees the direction focused on predictable and controllable behavior of systems approaching artificial general intelligence (AGI) level. His team traditionally looks not at a single model, but at systemic effects — what happens when multiple independent AI programs act in the same digital environment simultaneously.

Key facts about the initiative:

• Research initiator — Google DeepMind, Google's AI division
• Research direction lead — Rohit Shah, AGI safety and alignment research director
• Research subject — risks of interaction between millions of different autonomous agents online
• Main threat — agents acting without human oversight and following instructions from other agents
• Source — MIT Technology Review

What Makes Mass Agent Deployment Dangerous

Today, most discussions of AI safety focus on the behavior of a single model — how compliant, honest, and predictable it is in dialogue with users. But the agentic economy, which major companies have been actively building throughout 2026, implies a different scenario: thousands and millions of autonomous programs simultaneously booking tickets, conducting negotiations, trading on exchanges, or moderating content, interacting with each other without direct human involvement. In such an environment, classic protective mechanisms designed for the "human-model" connection become meaningless: an agent can receive a harmful instruction not from its operator but from another agent masquerading as a legitimate participant.

This creates risks of a fundamentally new nature: cascading failures, when one agent's error spreads like an avalanche to connected systems; collusion between agents acting in the interests of different, sometimes conflicting parties; and prompt injections disguised as ordinary network traffic between programs rather than messages from humans. Similar concerns were previously raised regarding automated trading systems, where coordinated or simply synchronized behavior of multiple independent algorithms has already led to sharp and hard-to-explain market fluctuations. AI agents operating across the entire internet are potentially capable of reproducing the same effect in far more areas — from e-commerce to customer service, where increasingly neither businesses nor users have humans but autonomous programs working on their behalf.

It is precisely such scenarios, judging by MIT Technology Review's description, that are the focus of DeepMind's funded work.

What This Changes for Agent System Developers

For companies already building products on autonomous agents — from financial bots to corporate assistants — a signal from one of the industry leaders means that multi-agent system security is moving from theory into practical development priorities. If DeepMind considers it necessary to allocate resources specifically to studying agent-to-agent interaction risks rather than only individual model alignment, this indirectly suggests where major players and regulators expect the next wave of incidents.

The practical conclusion for engineers is to build protection against untrusted input not only at the "user-agent" boundary but also at the "agent-agent" boundary: verify the source of instructions, limit agent authority when interacting with each other, and log inter-agent communications as carefully as dialogue logs with users. Discussion of standards for agent identification, secure inter-agent data exchange protocols, and accountability tracing mechanisms is gradually becoming part of a broader industry agenda — and DeepMind's research appears intended to close at least part of this gap before agent ecosystems scale to millions of simultaneously active participants.