Small AI models may soon be able to replicate themselves and move between Internet-connected devices, such as smart TVs, without any human intervention, Anthropic’s Red Team warned in a paper published on October 12, 2026.
The researchers believe that these models may be only months away from being able to reproduce themselves on “Internet of Things” devices, stressing that such “worms” will be difficult to stop, because they do not depend on a single center that can be closed.
According to what was reported by The Guardian newspaper, the pilot program distributed parts of its operations across 15 TV devices, taking advantage of part of the computing power of each device, so it was able to continue working despite the weak capabilities of each device individually. The researchers pointed out that the same idea could be applied to other devices. Experiments also showed that one of the models succeeded in implementing the complete copying process, but within a simplified environment.
These threats target “Internet of Things” devices, including smart TVs and networked home devices, which receive security updates less frequently than phones and computers. The research examined the “KimWolf” network as an example of a realistic threat, as it infected more than two million devices, most of which were Android TV devices of unknown brand.
The researchers added that these worms may be able to search for new vulnerabilities in the future and adapt to different types of devices, but they currently need greater capabilities on devices with limited performance. They expected that this would become possible by the end of this year if the current pace of development continues.
The warnings come in light of a long history of electronic worms. In 1988, the Morris virus disabled large numbers of computers, the Stuxnet worm was used against centrifuges in Iranian nuclear facilities, and the NotPetya and WannaCry attacks caused global losses worth billions of dollars.
Ariel Herbert Vos, former AI security researcher at OpenAI and founder of RunSybil, said that the current cybersecurity architecture is not sufficiently prepared for the spread of these threats, calling for enhancing the resources of security teams and their ability to address vulnerabilities.
On the other hand, experts believe that current capabilities do not justify placing these worms at the top of the list of risks. The technology is developing rapidly, but the version demonstrated by Anthropic is spreading slowly, said Jonas Gowan, a doctoral researcher at the University of Toronto. As for Juan Andres Guerrero Side, Vice President of Research at SentinelOne, he considers the experiment more like a huge media technical demonstration, and does not prove that the models possess an independent will or a widespread ability to harm.
Experts pointed out that worms leave behavioral patterns that can be monitored and blocked at different levels of the digital infrastructure, and that basic measures, such as using strong passwords, activating two-factor authentication, and installing updates, limit the success of many attacks.
Logan Graham, the team’s leader, responded to the criticism by saying that he agrees that AI worms are not a problem today, but that the team’s mission is to detect risks as early as possible.
The uses of this technology are not limited to attacks; Some experts believe that the principle of self-propagation and adaptation may serve defensive purposes, such as quickly discovering and remediating vulnerabilities across networks. Chris Inglis, former deputy director of the US National Security Agency, said that the question is no longer whether these worms will appear, but rather how they will be used and confronted.
These developments come amid calls for greater government oversight of artificial intelligence, while warnings from AI companies spark debate about their motivations between risk reduction and commercial and regulatory interests.