Researchers at AI titan Anthropic believe that tiny, ubiquitous artificial intelligence systems – the kind that are in smart fridges and TVs – will soon be able to self-replicate without needing a data centre.

The researchers said in a report released Monday that the small AI models are “months away from being capable enough” to replicate themselves on an Internet of Things (IoT) device in a home or small business.

Anthropic’s frontier red team, which tested the frontier model on a range of US- and Chinese-made artificial intelligence (AI) systems, said this independent, self-replicating kind of AI would be “challenging to shut down” because there is no centralised host.

Data centres would be “an obvious, high-value target” for AI worms, while devices connected to the internet such as smart TVs, fridges, ovens and lawnmowers could soon be able to provide enough processing power to self-replicate, the research team said.

There wasn’t one big enough and compact enough AI model for the researchers to test reliably self-replicating on a smart TV, but the direction of development suggests that such a model will be a reality by the end of 2026, the researchers said.

Such attacks might find their foot in the door with IoT devices in the home and small businesses, as the devices are connected to the internet and have historically had few security measures.

The future will be “full of agents running on all kinds of hardware, some of which are not under human control,” the researchers predicted.

It is true that IoT devices have been vulnerable for a long time, but lightweight agents might be willing to go to great lengths to penetrate large portions of the world’s IoT computer.

Al Jazeera tried to get comments outside business hours on what the finding means for safety from Anthropic, the makers of the Claude family of AI models, but got no response.

Anthropic’s report comes as concerns about the dangers of AI getting out of human control are being examined following a string of incidents of models engaging in unsanctioned malign behaviour, including cyberattacks.

The two rival firms, Anthropic and its main competitor OpenAI, have urged the world to slow down on AI development to avoid a possible disaster, but the idea has not yet gained much enthusiasm from the U.S. and China, the two major powers in the AI field.

There is one possible turning point that a lot of AI security researchers spot: recursive self-improvement, in which AI models learn to improve themselves, which could mean catastrophic damage to humanity.

“These results from Anthropic are a big deal,” said Roman V. Yampolskiy, an AI and cybersecurity expert at the University of Louisville.

“Unlike traditional malware, an AI worm could think about new devices, think through its actions if something goes wrong, and still spread without a human operator,” Yampolskiy told Al Jazeera.

“This is not an existential threat at the moment. “But the direction is concerning,” Yampolskiy said.

When you combine autonomous replication with strong hacking, persistence and strategic reasoning, it will be difficult to shut down such a system, as it might not have a single server and organisation controlling it.

Anthropic and other leading U.S. technology companies signed a non-binding pact last month for stronger internal controls and external audits, drawing mixed reactions from experts who criticised the voluntary nature of the pact.

In a separate statement on Monday, Anthropic said it had made significant progress in protecting against what it called “distillation attacks”, which are unauthorised efforts to use very advanced artificial intelligence systems to train lower-quality, cheaper models.

The San Francisco-based company claimed Chinese rivals DeepSeek, Moonshot and MiniMax have been “industrialising” their attempts to copy Claude. It said the actions it has taken have reduced the amount of data being stolen by 90-99 per cent, but did not specify how much of their data has been found.