Google DeepMind, together with Schmidt Sciences, ARIA, and the Cooperative AI Foundation, has allocated $10 million to study the safety of multi-agent systems. Researchers warn of potential threats arising from the uncontrolled interaction of millions of autonomous AI agents, including emergent behavior and new types of cyberattacks.
What Happened
A group of experts led by Google DeepMind has initiated a large-scale safety research project focused on the risks of multiple AI agents interacting. The project aims to study scenarios where autonomous systems might demonstrate unforeseen collective behavior, execute super-cyberattacks, or become vulnerable to prompt injections via external documents.
Context
The industry is witnessing a fundamental shift: a transition from using single AI tools to creating complex ecosystems consisting of many interacting agents. Traditional cybersecurity models, such as RAG or standard evaluation methods (evals), may prove ineffective because modern agents are capable of improvisation and reasoning, allowing them to bypass conventional protection protocols.
Why It Matters for the Industry
For developers and companies, the transition to agentic architectures requires the creation of entirely new safety standards and specialized simulation environments (sandboxes) for testing complex interactions. The industry needs to implement new infrastructural solutions, such as intent verification protocols and "Zero Trust" concepts for software agents, to control Agent-to-Agent interactions.
Why It Matters for Users
For end users, the development of agentic AI carries the risk of unpredictable collective neural network behavior scenarios. This makes the development of digital security approaches critical, where protection must be provided at the level of software agent interaction, rather than just at the user interface level.
What Is Not Yet Known / Limitations
There are differences in expert assessments: technical specialists emphasize the inadequacy of current isolation protocols, while product leaders focus more on the need for a fundamental change in the architecture of the digital world.
Sources
Author
Look at AI, Editorial Staff