Until now, hackers have used artificial intelligence as a tool they control from afar. Two independent research teams have now shown that the model itself could become the vehicle of an attack: a worm able to copy itself and spread across connected machines and devices without an operator guiding every step.

Psst!Add AlexTech.ai asPreferred Source on Google

From simple worms to adaptive ones

A computer worm is malware that spreads through a network by copying itself onto every system it can breach. Hackers can then use the compromised systems to steal data or link them into a botnet, useful for sending spam or knocking websites offline with fake traffic.

So far, worms have been small, simple programs built to exploit one specific security weakness. If that flaw was patched, the worm could no longer spread. An AI model directing its actions changes this: the worm can learn on the fly and choose its way in based on the device it finds. Jonas Guan, an AI researcher at the University of Toronto, says such a worm could be "much harder to stop".

The octopus worm

Guan and colleagues tested the prototype in a virtual network linked to several computers with GPUs, the chips used for heavy calculations. The simulated network also included fake office printers and security cameras. The worm's AI "head" started on a single machine and sent out "tentacles" to take control of smaller devices. When an attempt failed, it used what it had learned from the failure to try again. Whenever it found a computer with a powerful enough GPU, it copied its head onto that machine. The team also let the worm search a simulated internet for newly discovered vulnerabilities, while the real network stayed isolated to keep the test inside the lab.

According to Riegler of the Simula Metropolitan Center for Digital Engineering in Oslo, this ability to adapt to each target is "a new, hypercharged way of attacking systems". The most delicate finding is that the worm also found a way to survive a reboot of the infected device, something the researchers had not planned.

Limits and timing

The test conditions made spreading easier than it would be in the real world, Riegler noted; he was not involved in either study. The biggest practical obstacle is size. The Mirai worm, which took down sites including Reddit and Spotify in 2016, was a few kilobytes. The AI models used in the experiments weigh tens of gigabytes. Most common devices, laptops included, lack the space or the chips to run them.

New devices, such as some security cameras, are however starting to ship with hardware designed for AI. Anthropic argues the risks could materialize within months, but both Guan and Riegler think it will take longer. Riegler sums it up: "The ingredients are converging", though it is unclear how fast.