Anthropic researchers were able to run an AI model on 15 inexpensive set-top boxes. The experiment showed that powerful servers are not necessarily required for artificial intelligence to work. Scientists warn that AI viruses capable of independently spreading between smart devices may appear in a few months.
Anthropic researchers were able to run an AI model on 15 inexpensive set-top boxes. The experiment showed that powerful servers are not necessarily required for artificial intelligence to work. Scientists warn that AI viruses capable of independently spreading between smart devices may appear in a few months.
This is stated in the Anthropic study.
For the first experiment, the team purchased regular Android consoles for about $55 each. Each one was too weak to run a large AI model on its own, so the researchers distributed it across all the devices. And it worked. However, the model generated less than one token per second, meaning it took much longer to respond than a typical AI bot.
But the researchers were most interested in something else: could the AI independently search for new devices and launch on them? The researchers tested this in subsequent experiments.
First, the small Qwen3.5-9B model was tasked with independently configuring the launch of another neural network on different hardware. It coped with eight of the nine types of devices. Then Anthropic created an isolated network of eight virtual devices on Android. The AI's task was to find new devices, transfer the necessary files to them and configure the model's operation. GLM-5.2 did the best. In four out of five attempts, it managed to extend the AI to all eight devices.
However, this was not an AI hacking of devices, as the researchers had left the devices accessible without password verification beforehand. But even under these conditions, the experiment showed why the emergence of AI viruses worries researchers. The main difference between AI and viruses is that an AI model can adapt to new conditions. Instead of acting only according to pre-written instructions, it can potentially independently understand an unfamiliar device, fix errors and find a way to run on it.
Previously, dev.ua wrote that Anthropic CEO Dario Amodei warned about the risk of the emergence of uncontrolled groups of AI agents, which could pose a serious threat to the Internet in the next 6–12 months.
