Imagine an artificial intelligence capable of finding a computer connected to the internet, installing a copy of itself, and continuing to search for other devices to repeat the process. You know the idea of simply "turning off" an AI if something goes wrong? Well, tell me this: how do you shut down a system that no longer relies on a single machine to function? That is the possibility behind a recent warning from
Anthropic, the company responsible for Claude.
Instead of just investigating what AI models are currently capable of answering, a group of researchers began testing whether these models could acquire the skills necessary to spread across networks of devices. And there is a detail that makes this story even more unsettling: we aren't necessarily talking about massive data centers. In one of the experiments, the researchers distributed the execution of a 26-billion-parameter model across 15 small Android TV devices, each costing approximately $55.
The result was slow—less than one word per second—but it demonstrated that an AI can function by splitting its processing tasks among quite modest hardware. This does not mean the model hacked into or compromised these devices—they were specifically prepared for the test—but the demonstration changes how we envision the infrastructure required to keep an artificial intelligence operational.
The team then investigated another aspect of the problem: the ability to adapt programs to different types of machines. A model with approximately 9 billion parameters managed to set up execution environments on eight out of the nine types of hardware tested. This wasn't just a matter of copying a file; it involved navigating differences between systems and finding ways to get the software running. Finally, they moved to the scenario that prompted the warning: in an isolated network of eight virtual machines, one agent was tasked with establishing new installations.
One of the models tested managed to reach eight devices in four out of five attempts. This is where the story takes on a new dimension: malicious software capable of spreading across networks—known as "worms"—has existed for decades. The concern is that AI agents might add something to this process that traditional programs lack: the ability to interpret obstacles and improvise solutions.
Imagine a copy of such a program encountering a device with different configurations than those it has seen before; instead of simply failing, it could try a different approach, adjust its commands, and continue the operation. Researchers also observed that these agents could benefit from information left behind by previous attempts—such as a log of errors encountered and the solutions that proved successful. However, there is a crucial distinction between these results and the idea of an AI spreading freely across the internet. None of these experiments involved a model invading actual consumer devices and creating independent copies of itself on a global scale; the tests took place under controlled conditions, with facilitated access and, in part, the use of external computing resources.
Anthropic is raising concerns that these capabilities—demonstrated individually or in simplified environments—could evolve into a more autonomous propagation system. If that were to happen, the issue would not just be an AI performing tasks without supervision; it would be an AI capable of maintaining its operations across devices scattered throughout various networks, where shutting down a single server might no longer be enough. Significant barriers still exist, such as limited processing power in inexpensive devices, the need for access to the hardware, detectable network traffic, and security systems that can halt such propagation. Therefore, we are not dealing with an artificial intelligence that has escaped control—at least, not yet. Rather, this is research aimed at understanding whether some of the barriers to such a scenario are beginning to diminish.
