As artificial intelligence (AI) advances, internet-connected devices that exchange data with one another, in the so-called “Internet of Things” (IoT), could become a problem.

According to an Anthropic report released in October, AI-powered malicious software could install itself on this type of device. And spread autonomously.

AI ‘worms’ could spread like ‘digital mould’, Anthropic warns

Computer viruses that spread from one machine to another on their own are known as “worms”.

They follow a pre-programmed script, exploiting known security vulnerabilities or poorly protected access points.

The fix is usually to patch flaws and close security gaps. With AI “worms”, however, things get more complicated. And more dangerous.

That is because this more advanced type of “worm” would be able to solve problems autonomously and adapt to different kinds of devices.

Researchers at Anthropic – the developer of Claude – showed that AI applications can run without the need for a large data centre.

In a controlled test environment, the researchers created simulations of eight Android devices.

The AI model tested installed itself on one of these devices, replicated from it and infected the others.

In four out of five attempts, the model reached all eight targets. In the other, it infected seven.

If an AI model of this kind were combined with malicious software, the “worm” could learn to get around security measures. And it would take roughly one to two days to do so, according to Anthropic’s developers.

An army of infected devices

On top of that, the AI could link the infected devices into a botnet.

A single infected device has little processing power. But the more devices are infected, the faster the “worm” could spread.

Connected devices could be taken off the network. And security flaws could be patched.

However, the decentralised structure of this type of malware would make it very hard to contain.

Anthropic describes this kind of infection as a slowly growing digital slime mould.

The company believes smaller AI models could reach this level of spreading capability as early as 2026.

(The Anthropic report is available on the company’s website. This article also used information from DW.)