Google DeepMind funds research into the risks of millions of AI agents interacting

Google DeepMind is funding new research into the risks that arise when millions of different AI agents interact with one another on the internet. MIT Technology Review reported the news, citing Rohin Shah, who heads AGI safety and alignment research at the company. According to him, the mass rollout of agents that can perform tasks without constant human control and receive instructions from other agents creates a fundamentally new class of risks that the industry has so far studied only weakly.

Who is behind the research

Google DeepMind is Google's research division, responsible both for developing frontier models in the Gemini line and for fundamental AI safety research. Rohin Shah oversees the area that deals with the predictable and controllable behavior of systems approaching the level of artificial general intelligence. His team traditionally looks not at a single model but at system-level effects: what happens when many independent AI programs act at the same time in a shared digital environment.

Key facts about the initiative:

• Initiator of the research: Google DeepMind, Google's AI division
• Lead of the area: Rohin Shah, head of AGI safety and alignment research
• Subject of study: risks of millions of different autonomous agents interacting online
• Main threat: agents that act without human oversight and obey instructions from other agents
• Source: MIT Technology Review

Why mass deployment of agents is dangerous

Today, most discussions of AI safety focus on the behavior of a single model: how obedient, honest and predictable it is in a conversation with a user. But the agent economy that major companies have been actively building throughout 2026 implies a different scenario: thousands or millions of autonomous programs simultaneously booking tickets, negotiating, trading on exchanges or moderating content, coming into contact with one another without direct human involvement. In such an environment, classic safeguards designed for a "human–model" pairing lose their meaning: an agent can receive a malicious instruction not from its operator but from another agent posing as a legitimate participant in the process.

This gives rise to risks of a fundamentally new nature: cascading failures, in which 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 as a message from a human. Similar concerns were raised earlier about automated trading systems, where the coordinated, or simply synchronized, behavior of many independent algorithms has already caused sharp and hard-to-explain market swings. AI agents operating at the scale of the entire internet could potentially reproduce the same effect in far more areas, from e-commerce to customer service, where increasingly an autonomous program, not a human, works on the side of both the business and the user.

According to MIT Technology Review's description, these are the scenarios at the center of the work DeepMind is funding.

What this changes for developers of agent systems

For companies already building products on autonomous agents, from financial bots to corporate assistants, the signal from one of the industry's leaders means that multi-agent safety is moving from the theoretical plane to the practical development agenda. If DeepMind considers it necessary to devote resources to the risks of interaction between agents, and not only to the alignment of individual models, that indirectly points to where major players and regulators expect the next wave of incidents.

The practical takeaway 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 agents' privileges when they interact with each other, and log inter-agent communications as carefully as logs of conversations with users. Discussion of agent identification standards, protocols for secure inter-agent data exchange and mechanisms for tracing accountability is gradually becoming part of the industry's broader agenda. The DeepMind research, by all appearances, is meant to close at least part of this gap before agent ecosystems scale to millions of simultaneously acting participants.