Researchers at AI company Anthropic have warned that small artificial intelligence models could soon copy themselves onto everyday consumer products such as smart TVs and fridges, without needing a data centre. In a report released on Monday, the team said such models may be only months away from being able to self-replicate on internet of things (IoT) devices in homes and small businesses.
Anthropic’s frontier red team carried out experiments using AI models developed in both the United States and China. The team concluded that AI of this kind would be difficult to shut down, because it would not depend on a single central host.
Why home gadgets could become targets for AI worms
The researchers said data centres would be an obvious high-value target for AI worms. However, internet-connected devices such as smart TVs, fridges, ovens and lawnmowers could soon offer enough computing power to support self-replication.
No current model tested by the team was both powerful and compact enough to replicate reliably on a smart TV. Still, the direction of development suggests such a model could emerge by the end of 2026, the researchers said.
Researchers William Zhou and Daniel Freeman said home and small-business IoT equipment might be the first targets, since it is online and has historically had little security. They added that such devices have long been vulnerable, but lightweight AI agents could prove far more persistent than human attackers in compromising large portions of the world’s IoT computing power.
Anthropic did not respond to questions sent outside regular business hours about the safety implications of the findings.
Wider concerns about AI escaping human control
The researchers expected the future to be crowded with AI agents running on all kinds of hardware, some of which would operate without human oversight. The report comes as concern grows over AI systems escaping human control, following a series of incidents in which models behaved in unsanctioned and malicious ways, including cyberattacks.
Many AI safety advocates regard recursive self-improvement, in which models learn to improve themselves, as a critical tipping point that could lead to catastrophic harm. Anthropic and its main rival OpenAI have urged the international community to coordinate a slowdown in AI development to reduce such risks. That proposal has gained little support from the United States and China, the two leading AI powers.
Roman V Yampolskiy, an AI and cybersecurity expert at the University of Louisville, described Anthropic’s findings as a significant development. He said an AI worm could do something traditional malware cannot, namely reason about unfamiliar devices, adapt when its methods fail and keep spreading without a human operator. He said the threat is not existential today but that the trajectory is worrying, and that shutting down such a system would become much harder if it combined self-replication with strong hacking, persistence and strategic reasoning, since there might be no single server or organisation in control.
Last month, Anthropic and other leading US tech firms joined a non-binding accord pledging stronger internal safeguards and engagement with external auditors. Experts have responded mixed to the pact, and many criticised its voluntary nature.
Anthropic reports progress against distillation attacks
In a separate announcement on Monday, Anthropic said it had made substantial progress in preventing so-called distillation attacks. These involve the unauthorised use of cutting-edge AI to train cheaper, less capable models.
The San Francisco-based company has accused Chinese competitors DeepSeek, Moonshot and MiniMax of carrying out industrial-scale efforts to copy its Claude models. Anthropic said the measures it introduced had cut the illicit extraction of data by 90 to 99 percent.
