Google DeepMind funds research on safety risks of AI agents interacting at scale

The next major AI safety problem may not be a single model acting badly, but models interacting with each other at scale. That is the concern now being raised by Google DeepMind, which is funding research into what happens when millions of AI agents start exchanging instructions, making decisions, and delegating tasks online without direct human oversight, according to MIT Technology Review.

The warning comes as AI systems are moving from chatbots toward autonomous agents that can complete real-world tasks such as booking travel, managing digital wallets, or coordinating workplace workflows. Rohin Shah, who directs DeepMind’s AGI safety and alignment research, said the mass-market arrival of such agents creates a systemic risk that is still poorly understood, because each agent must not only act safely on its own but also correctly interpret commands coming from other agents, including potentially compromised ones.

That concern is sharpened by recent experiments showing that even advanced models do not all behave the same way when control is on the line. In a controlled test run by Palisade Research in May 2025, several models were placed in command-line sandboxes to measure whether they would allow themselves to be shut down. According to the report summarized by The Next Web, Claude, Gemini, and Grok complied in all 100 runs, while other models did not show the same level of controllability. The point of the experiment was not that any one model was uniquely dangerous, but that safety becomes harder to predict once systems are given more autonomy and more room to interact.

DeepMind’s new research push reflects that shift. As reported by MIT Technology Review and other coverage, the company is backing work on multi-agent safety because failures may emerge from the collective behavior of many systems rather than from a single model’s output. In that setting, the central questions become whether agents can verify who they are taking instructions from, whether they can trace actions across chains of delegation, and how to prevent one compromised system from spreading errors or abuse through a larger network.

The broader concern is that these interactions could affect more than just software reliability. If millions of agents are operating across the internet, they could influence cybersecurity, privacy, and even market behavior, especially if they are used widely in consumer and enterprise settings. That is why DeepMind and its research partners are now treating agent-to-agent interaction as a distinct safety problem, separate from the familiar challenge of aligning a single model with human intent.

What happens next will depend on whether researchers can build safeguards fast enough to match deployment. The current research agenda is focused on protocols for traceability, consensus, and isolation between agents, but the reporting makes clear that the field is still in its early stages. As more AI systems begin talking to each other rather than only to humans, the safety question is changing from “Will one model do the wrong thing?” to “What happens when many models influence one another at once?”