Plate Nº 42 · recorded November 16, 2026

Reported finding

Two teams show how AI worms could spread through smart devices

Two labs turned AI models into self-copying worms that adapt to different devices. Both ran in lab settings, and real obstacles remain before anything reaches your refrigerator.

By James Calloway5 min read991 words

In brief

  1. Anthropic researchers ran a 'digital slime mold' AI worm across 15 networked Android smart TVs, in a report posted online October 12
  2. The octopus-worm team's preprint appeared on arXiv.org on June 2 and is now in peer review
  3. Mirai was a few kilobytes in size; the medium AI models in the experiments were tens of gigabytes
  4. Anthropic puts the AI-worm timeline at months away, while Riegler and Guan say it will probably take longer
  5. Mirai infected hundreds of thousands of routers, cameras and baby monitors and knocked Reddit and Spotify offline in 2016
We know how to stop computer worms. But what if AI helps them persist?
Plate Nº 42We know how to stop computer worms. But what if AI helps them persist? — vizdomtech / pixabay

Two independent research teams have shown that an AI model can be turned into a worm — malware that copies itself onto machines it hijacks, without a human steering each step. Neither team let anything loose on the real internet, but one built a worm that survived a reboot, and that was not part of the plan.

How a worm works, and why AI changes it

A worm squirms through the internet, copying itself onto whatever systems it can take over. The attacker then mines those systems for data, or links them into a botnet and uses the pooled computing power to send spam or flood websites with traffic until they fall over.

Until now, worms have been small, fixed programs, preprogrammed to exploit one specific weakness. Patch that hole and the worm dies. Point an AI model at the job and the worm can change tactics mid-flight, picking its attack based on the device in front of it, says Jonas Guan of the University of Toronto.

The octopus: one brain, many tentacles

Guan and colleagues built their creation in a virtual network anchored by several computers with graphics processing units, or GPUs — chips built for heavy math. The simulated network also held fake office printers and security cameras. An attacker would probably launch a worm like this from the cloud.

  • The head started on one GPU machine, then sent tentacles out to take control of smaller devices.
  • Facing a new device, the head used AI to plan how a tentacle might get in. Many attempts failed, and the octopus used what it had learned from each failure to invent another.
  • Whenever it found a computer with a large enough GPU, it copied its head there.
  • Devices too small to host the model still served as way stations for finding new hosts.

Michael Alexander Riegler, a cybersecurity researcher at the Simula Metropolitan Center for Digital Engineering in Oslo who had no part in either study, calls that failure-driven persistence "a new, hypercharged way of attacking systems."

The head also searched a simulated internet for information about newly discovered vulnerabilities. The team never connected it to the real one, and spent considerable effort making sure it could not escape the lab, Guan says. The team posted a preprint on arXiv.org on June 2, and it is now in peer review. Then a surprise: the octopus-worm came up with a way to survive a reboot, which Guan says they had never asked it to do.

The slime mold: thinking spread across devices

Anthropic researchers took a different route, building on Guan's work. They called theirs a "digital slime mold" because it breaks itself into pieces and runs those pieces on many small devices: 15 networked Android smart TVs in their report, posted online October 12.

Anthropic declined to comment for this story, but Guan says the company's work shows "unlike human reasoning, AI reasoning can be distributed." The chopped-up model processes data thousands of times more slowly than one sitting inside GPUs, because data has to travel over internet connections, possibly around the world, rather than staying on chips optimized for AI.

That pace was too slow to prove worm-like spread, so the Anthropic team ran a separate GPU experiment that did show spreading and adaptation to different kinds of devices.

What stands between this and your kitchen

Riegler notes that both test environments made the worms' job easier than it would be in the real world. Size is another wall: Mirai, which knocked down Reddit and Spotify in 2016, was a few kilobytes, while the medium AI models in these experiments ran tens of gigabytes. The top-tier models behind the "rogue AI" incidents that made headlines earlier in 2026 were bigger still, and were not worms, because they never copied themselves and spread.

Most devices, typical laptops included, lack the space and the chips to run worm-capable models, Guan says. But new hardware is arriving: security cameras now ship with advanced hardware for running AI, Anthropic researchers write, and smaller models grow more capable of sophisticated cyberattacks.

On timing, Anthropic puts the threat at months into the future; Riegler and Guan say it will probably take longer. "The ingredients are converging," Riegler says.

Could this reach my devices?

Not soon, and not the way Mirai did. Mirai went in with a small set of common default usernames and passwords, then infected hundreds of thousands of home routers, security cameras and baby monitors.

Zane Ma of Oregon State University in Corvallis, who presented an analysis at a 2017 security symposium, said "just using this very simple technique, this Mirai botnet was able to infect hundreds of thousands of devices." Reddit's outage was collateral damage from a small group of young hackers who wanted to mess with gaming servers.

A reboot cleared Mirai, and changing default credentials blocked the same route. An AI worm would make cleanup harder, because it adapts.

What could stop an AI worm

  • Its sheer bulk. Big models guzzle computing power, and that draw shows up on a network.
  • Ordinary hygiene. Strong passwords, regular updates and other standard security measures would have blocked both attacks demonstrated in the experiments, Riegler says.
  • Provider action. Ma notes your internet service provider would realize "there's something weird going on on these devices" if a refrigerator or microwave began pushing heavy traffic, then quarantine it by cutting it off.

Home routers act as a central hub for most smart devices, which adds a layer of protection, and Ma points out that most home gadgets never touch the public internet directly.

Guan says defenders should get ready regardless of the timeline, and that he and colleagues are already testing specific countermeasures, including an AI-driven defender. Trying to anticipate emerging threats and protect systems ahead of time is "what I'm passionate about," he says. "We are working on the problem day and night."

via sciencenews.org (Source)

Filed under

  • ai-worms
  • artificial-intelligence
  • cybersecurity
  • malware
  • smart-devices
Share this article:

More from James Calloway

James Calloway

Show full bio

Staff writer covering regulation, policy and their impact on general science news at SciBeat.

411 articles

Nearby plates

« Previous articleNext article »