AI is already powering new, more adaptable forms of cyberattack, but hackers typically wield the technology from afar. What if they set the AI models themselves loose to self-replicate and spread in a malicious manner?
Recent experiments have demonstrated two different ways this could happen. Two independent teams showed how a cyberattacker could turn an AI model into a “worm,” a type of malware that, unlike a computer virus, spreads on its own. And AI unlocked a new worm power: adapting attacks to different devices.
The results hint at a new way AI could be used to ramp up cybersecurity threats — not just to computer systems but to smart devices such as security cameras or refrigerators. The good news is that the world still has time to build up defenses as several obstacles remain before this type of malware would work in the real world, says AI researcher Jonas Guan.
AI could indeed make a worm “much harder to stop,” he says, but “there are a lot of kinks to iron out before something like this really works at scale.”
Morphing simple worms into shrewd attackers
Here is how a worm cyberattack works. A malware worm squirms through the internet, copying itself onto whatever systems it can hijack. Hackers can then mine hijacked systems directly for their data or link them together into a botnet and use its computing power, often to send spam email or overwhelm websites with traffic, knocking them offline.
Historically, computer worms have been small, simple programs preprogrammed to take advantage of specific security weaknesses. So they could only get into devices that had that weakness. When defenders patched a worm’s target, it could no longer spread. With an AI model directing its actions, however, a worm can adapt and learn on the fly, choosing its line of attack depending on the device, says Guan, of the University of Toronto.
That adaptability enables the worm to burrow into many different types of devices, both experiments showed. When hackers unleash AI from afar, they orchestrate what the model does. An AI worm that copies itself widely could hypothetically escape anyone’s control.
South_agency/E+/Getty Images
Still, there’s a caveat to the experiments’ results, says Michael Alexander Riegler, a cybersecurity researcher at Simula Metropolitan Center for Digital Engineering in Oslo, Norway. Both teams’ test environments made it easier for the AI worm to spread than it would be in the real world, says Riegler, who was not involved in either study.
And there’s another practical obstacle to the unleashing of AI worms. They use more computing power and take up more space than a conventional worm. The Mirai worm, which took down a slew of websites including Reddit and Spotify in 2016, was a few kilobytes in size, while the medium-sized AI models in the experiments were tens of gigabytes. That’s still way smaller than the massive, top-of-the-line models behind the “rogue AI” incidents that made headlines earlier in 2026. (And those weren’t worms, because they didn’t copy themselves and spread.)
Most devices — including typical laptop computers — don’t have enough space or the right kind of chips to run worm-capable AI models, Guan says. But new versions of devices, such as security cameras, are coming out with advanced hardware for running AI, researchers at the AI company Anthropic write. And at the same time, even smaller AI models are becoming more capable of sophisticated cyberattacks. These developments could expand the pool of devices an AI worm could exploit.
“The ingredients are converging” to make AI worms more realistic, Riegler says. How quickly they will converge is up for debate. Anthropic puts the timeline at months into the future, but both Riegler and Guan say it will probably take longer.
Spreading like octopuses or slime molds
The Mirai worm spread itself through smart devices including home routers, security cameras and even baby monitors. It broke into these devices by trying a small set of common default usernames and passwords. “Just using this very simple technique, this Mirai botnet was able to infect hundreds of thousands of devices,” says cybersecurity expert Zane Ma of Oregon State University in Corvallis, who presented an analysis of the attack at a 2017 security symposium. High-profile downed sites such as Reddit were actually collateral damage: A small group of young hackers had created the worm to mess with gaming servers.
Rebooting an infected device cleared Mirai. Changing its default login credentials prevented reinfection through the same route. But an AI-powered worm might make cleaning up this type of cybersecurity breach much more difficult. Why? Because the AI worm acts more like an octopus, Guan says.
Guan and colleagues tested this octopus-worm in a virtual network. Anchored by several powerful computers that each had a chip called a graphics processing unit, or GPU, the faux networks also contained simulations of smart devices, including office printers and security cameras.
The octopus’s AI model head started on a single machine with a GPU. From there it sent out tentacles to take control of smaller devices. Upon encountering a new device, the head would use AI to plan how its tentacle could hack in. Often, a hack wouldn’t work, but the octopus wouldn’t give up. It would use info about how it failed to come up with a new hack to try. That ability is “a new, hypercharged way of attacking systems,” Riegler says. Whenever it found a computer with a large enough GPU, it would copy its head onto that new machine.
The head could also search a simulated internet for information about newly discovered vulnerabilities. (The team didn’t connect it to the real internet for obvious reasons — protection against an AI octopus invasion — and spent considerable effort making sure their creation couldn’t escape the lab, Guan says.)
The octopus-worm even came up with a way to survive a reboot, something Guan says they hadn’t asked it to do. The team posted a preprint of their findings at arXiv.org on June 2, and the paper is now going through peer review.
Meanwhile, at Anthropic, researchers showed another way an AI worm might work by building on Guan and his colleagues’ research. If an octopus-worm wasn’t weird enough, Anthropic dubbed its creation a “digital slime mold”, a name that captures the model’s slow, oozing behavior. Anthropic declined to comment for this story, but Guan says the company’s work shows that “unlike human reasoning, AI reasoning can be distributed.” That means an AI worm wouldn’t necessarily need to hijack a powerful computer to copy and run itself — it could spread its “thinking” out among smaller devices.
And that’s exactly what the slime mold version of an AI worm does. It breaks itself up and runs pieces of the model on multiple smaller devices. In their report posted online October 12, the Anthropic researchers used 15 Android smart TVs that they had networked together.
Anthropic’s chopped-up model processes data thousands of times more slowly than what’s possible with GPUs similar to the ones Guan and colleagues relied on. That’s because the model’s data has to travel over internet connections that could take it all around the world during processing, rather than remaining in one place, within chips optimized for AI processing. That pace was too slow to demonstrate an ability to spread like a computer worm, Guan notes. So the Anthropic team ran a separate experiment with an AI worm running on GPUs that showed it could spread and adapt its attacks to different kinds of devices.
What could stop an AI worm?
Whether AI worms are coming in months or whether we have longer to prepare, defenders should get ready now, Guan says. And AI worms do have weaknesses.
For one thing, the large size of an AI worm could be its downfall. Even if smaller AI models become more capable, these are still big programs that guzzle computing power. You might not notice if your fridge or microwave suddenly started sending lots of traffic onto the network, but your internet service provider would. They’d realize “there’s something weird going on on these devices,” Ma says. From there, they could quarantine affected devices, disconnecting them from the internet to prevent further spread.
Ma also points out that most smart devices in a person’s home aren’t directly connected to the public internet. Most home networks use a router as a central hub, and that provides a layer of protection against a worm attack.
Strong passwords, regular updates and other normal security measures would also protect against the attacks demonstrated in both experiments, Riegler says.
Guan and some colleagues are now testing specific defenses to stop AI worms, including devising an AI-driven defender. Trying to anticipate emerging threats and protect systems ahead of time is “what I’m passionate about,” Guan says. “We are working on the problem day and night.”





