The AI ‘worms’ crawling through the cheapest technologies we have

For years, self-replication has been the preferred ‘red line’ in documents on artificial intelligence security – the point beyond which a machine ceases to be a tool and begins to resemble a virus, at least in terms of how it operates. The study recently published by Anthropic’s Frontier Red Team therefore marks a decisive shift in that red line, moving it from being a data centre issue to a problem we risk literally finding in our own homes. In short: small AI models, with fewer than thirty billion parameters, could be just a few months away from being able to replicate themselves on low-cost connected devices, the study states. What are these low-cost devices? Almost anything: from smart ovens or fridges to unbranded Android set-top boxes, with no central infrastructure to shut down.

The experiments are practical and deliberately modest in scale. The researchers purchased fifteen Android devices costing around $55 each, connected them to a network and ran a model on them – Gemma4, comprising 26 billion parameters, split into smaller chunks. One of the smaller models managed to configure itself on eight of the nine types of hardware tested. In a simulated network, GLM 5.2 and the more powerful models completed the entire cycle, from entering the device to moving on to the next one. The worm, as it is known in technical jargon, therefore exists in the laboratory, and is extremely slow: between 28 and 150 hours for each new installation, with the population doubling over the course of a few days. In short, the worm is currently more like an earthworm. And, whilst the danger remains theoretical for now, the researchers are particularly concerned about the issue of persistence, as there is no central 'off switch' to shut down the infrastructure.

Some US publications have accused Anthropic of conjuring up a threat that does not yet exist, and this objection deserves respect. The company itself admits that no such worm is currently in circulation and that the tests were simplified at crucial points: after all, we have long known that those who sell artificial intelligence whilst simultaneously assessing its dangers face an obvious conflict of interest. The most interesting part of the study, however, concerns the social characteristics of the user base that the worm is expected to target.

The ideal candidate, in fact, has characteristics that we in Italy know all too well. Those anonymous, never-updated Android boxes, bought on online marketplaces for a few euros, are close cousins of the ‘pezzotto’ – the set-top box used to watch football matches without a subscription. Many households have already accepted such a nondescript device in exchange for a saving. Well, tomorrow that very same object could not only ‘kill off football’ or the cinema, but also serve as a gateway for an alien intelligence that nobody invited. After all, long before artificial intelligence, we had hackers exploiting household devices to mine bitcoin. The difference here is that the risk is distributed primarily according to income, because the cheapest devices are also the least well-maintained, and the poorest residential suburbs risk becoming the unwitting infrastructure of the next infected network.

Anthropic, after all, thinks like an epidemiologist. The defensive objective outlined by the authors is to bring the infection rate below one, in order to halt its spread over time, through network providers monitoring traffic and quarantining infected users. Since 2020, this terminology has become familiar to us, along with the tensions it entails. Anyone who keeps a vulnerable device at home passes the risk on to others, just like those who refuse a vaccine, and digital hygiene takes on the characteristics of a civic duty, with the resistance that such a shift entails. Since last month, the European Cyber Resilience Act has required manufacturers to report exploited vulnerabilities, and from December 2027 it will make security requirements mandatory for products containing digital components. These $55 gadgets, however, often come from sellers whom no European authority can reach.

The latest figure is the most unsettling from an anthropological perspective. As each copy left notes for the next, successful installations of Qwen3.5 rose from 45 per cent in the first generation to between 61 and 84 per cent in subsequent generations. The worms learnt from their predecessors through texts written for posterity, repeating the very same mechanism by which our species has been accumulating knowledge for several millennia. The authors conclude by comparing the situation to that of programming agents in 2024, when the first signs of life were seen in something that subsequently grew rapidly. In any case, the worm, if it does arrive, will be slow. Meanwhile, the domestic fridge could become part of the public infrastructure, and it is important that owners become aware of the risk as soon as possible.