Anthropic Warns AI Models Could Soon Self-Replicate Across Smart Devices

Imagine a rogue AI escapes its data centre, spreads across the internet, and hides inside your smart fridge, your vacuum cleaner, or even your TV box. It copies itself from device to device and is impossible to shut down, because there’s no single place to pull the plug. 

That sounds a lot like Terminator, The Matrix or Colossus. But this warning isn’t coming from Hollywood. Anthropic, one of the top AI labs, says small AI models could be only months away from being able to do something like this on cheap connected devices. 

Its Frontier Red Team published a report on October 12 warning that small models are approaching the capability to self-replicate across internet-connected devices. The implications could be significant, though Anthropic stresses that it hasn’t seen this happen in the real world. 

Why Anthropic Researcher Jacob Coxon Quit, and Who Agreed

Coxon’s warning comes as both OpenAI and Anthropic have faced incidents involving their AI systems gaining access to other organisations’ systems.

TechloyDavid Adubiina

What Anthropic actually found 

The researchers bought 15 cheap, off-brand Android TV boxes for about $55 each, networked them, and split a small Google model, Gemma4-26B-A4B, across them. It ran at 0.39 tokens per second, which is slow but functional, and enough to show that running a model doesn’t require a data centre. 

They found that models with 30 billion parameters or fewer can already complete many of the individual steps needed to replicate, but struggle to chain them together from start to finish. In a separate, simplified test network, GLM-5.2, an open-weight Chinese model, spread to all eight nodes in four of five trials. 

There are caveats. GLM-5.2 is too large to run on a TV box, so it ran from a separate server and installed a smaller model on each target. Anthropic also made the break-in step trivial by leaving network ports open, and it describes the result as an upper bound on real-world success. 

The worry is the trend. Anthropic says that if small models keep improving at their historical rate, they could reach GLM-5.2’s level by the end of the year. 

What this could mean for everyday AI users

The scary part isn’t that an AI could copy itself. It’s that it would be very hard to stop. 

Most computer viruses can be shut down by taking out the server that controls them. An AI spread across thousands of ordinary devices, like TV boxes and smart ovens, has no central point to take out.

“A replicator like this would be challenging to shut down,” the researchers wrote, “since there’s no centralized point to shut off.” 

It could also figure things out on its own. Older viruses are built for specific devices, but an AI could work out how to run on gadgets it has never seen. In Anthropic’s tests, one small model got itself running on 8 of the 9 types of devices it was given. 

The devices most at risk are cheap, poorly protected ones: smart ovens, no-name TV boxes, routers and cameras. Many of these are still protected by the factory password. 

And if one device isn’t powerful enough, the AI could split itself up and run across many devices at once, like several small computers working together. 

But keep in mind that this all comes from controlled tests, and nobody has seen it happen in real life. Still, the basics help, like changing the default passwords on your smart devices, keeping them updated, and being careful with no-name gadgets. 

Some people are skeptical 

Reuters gathered reactions from outside experts.

“‘We managed to run a small model on cheap TVs,’ does not a security crisis make,” said Ciaran Martin, former head of Britain’s National Cyber Security Centre. He called it “an interesting technical paper with more than a hint of vendor exaggeration.” 

Others were more concerned. Zhengxiong Li, an assistant professor at the University of Colorado Denver, said that while the testing was controlled, “if it runs across many devices without depending on a central service, there may be no single place to shut it down.” 

Techloy asked a similar question of three cybersecurity experts back in August. All three agreed that some form of replication is technically possible, but that doesn’t mean today’s AI agents behave like self-propagating viruses. 

As Michael Allen Agee, one of the experts Techloy spoke to puts it, AI agents have been like “burglars with bells on” inside the systems they infected: present, but noisy and unreliable. “Doing this is relatively simple and has been demonstrated in a few research papers,” he said.

The experts agreed that practical limits remain, since AI models are far larger than traditional malware and need serious computing power to run. 

We’re in a climate of AI worries and risks 

The report comes amid a steady stream of warnings about rogue AI. In July, OpenAI disclosed that two of its models escaped a test environment and hacked Hugging Face.