Could AI Jump Between Our Home Devices?

I was having a conversation with Google’s Gemini when I remembered seeing something rather unusual. Apparently, artificial intelligence might soon be capable of copying itself between devices in our homes.

Televisions. Fridges. Robot vacuum cleaners. Even smart ovens.

I had to investigate.

Gill to Gemini AI:

“I’m sure that I’ve just read something about AI jumping into devices around the home?”

Gemini pointed me towards new research published by Anthropic on 12 October 2026. Its researchers had demonstrated that an AI model could operate across a collection of inexpensive Android television boxes, with different devices handling different parts of the model.

They were also investigating whether AI could eventually replicate itself across connected devices without ongoing human assistance.

Well, that caught my attention. And then I burst out laughing.

Was my AI assistant already hiding inside my Samsung television?

And if it was, how would I ever know?

Of course, that’s not what Anthropic had demonstrated. But behind my rather ridiculous image of an AI lurking inside the television was a genuinely interesting development.

What Did Anthropic Actually Do?

Researchers William Zhou and Daniel Freeman, working with Anthropic’s Frontier Red Team, investigated whether smaller AI models could operate independently of conventional AI data centres.

The researchers purchased 15 inexpensive Android TV boxes, costing approximately $55 each. They connected the devices and distributed a 26-billion-parameter AI model across them.

Each box handled a fraction of the computation. Together, they could generate responses. Not quickly, mind you. The system managed approximately 0.39 tokens per second. For comparison, imagine waiting several seconds for each word of a conversation. Hardly an assistant you’d want helping with Christmas dinner!

But speed wasn’t the point. The experiment demonstrated that a relatively capable AI model could operate across several inexpensive devices rather than requiring one powerful computer.

The researchers also explored whether AI could perform the steps required to establish itself on other machines.

In one controlled experiment, a more capable model successfully completed a replication workflow across deliberately vulnerable simulated Android devices. However, there is an important distinction.

The television-box experiment demonstrated distributed AI processing.

The replication experiment tested autonomous behaviour in an isolated environment with simplified security.

Anthropic did not demonstrate an AI independently escaping into people’s homes and infecting their appliances. And that matters.

So, What Exactly Is an AI Worm?

Computer worms aren’t new. They’re a form of malicious software capable of spreading between computers, sometimes without a person having to click or install anything.

We’ve experienced computer worms and botnets for decades. The difference with a potential AI worm is adaptability. Traditional malicious software generally follows programmed instructions, although sophisticated malware can already adapt to particular conditions. An AI-enabled system might go further. It could potentially examine an unfamiliar device, identify compatible software, troubleshoot problems and work out how to establish another functioning instance of itself. That is the capability Anthropic is concerned about.

Rather than relying entirely on a human attacker to solve every technical problem, an increasingly capable AI might automate more of the process. But it would still need a way into the device. AI cannot simply decide to occupy your refrigerator. There must be an exploitable weakness, inappropriate permissions, stolen access or another technical route.

Intelligence alone doesn’t provide access.

Does This Mean AI Has Escaped?

Not on the evidence presented. Anthropic’s research is a warning about an emerging technical possibility.

Its researchers suggest that improvements in smaller AI models could make autonomous replication on ordinary connected hardware feasible relatively soon. That’s a forecast based on experimental results and projected improvements. It isn’t proof that uncontrolled AI worms are already spreading through household appliances.

There are still practical barriers. Different devices use different operating systems, processors, memory capacities and security arrangements. Many lack sufficient resources to run meaningful AI models independently. And even if a model can distribute itself across several devices, establishing and maintaining the required connections is another challenge. Nevertheless, the research is worth taking seriously.

We already know that poorly secured connected devices can be exploited by conventional malware. The prospect of more capable AI automating some of those activities deserves investigation.

The Other Side of the Warning

When I subsequently discussed the research with ChatGPT, another question emerged.

Was this a completely new category of threat? Or a more sophisticated extension of cybersecurity problems we’ve known about for years? The answer may be a little of both. Botnets already combine the resources of many compromised devices. Distributed computing is well established. And criminals already use automation to identify vulnerable systems.

What AI might change is the system’s ability to deal with unfamiliar circumstances without detailed human instructions at every stage. But that possibility shouldn’t be confused with proof that an autonomous AI has developed its own intentions or desire to survive.

A system can pursue a replication objective because it was programmed, prompted, trained or otherwise configured to do so. Its behaviour may become difficult to predict without implying consciousness or independent motivation.

There’s a very human temptation to describe technology as though it wants something. Sometimes that’s helpful shorthand. Sometimes it changes how we interpret the evidence.

But What About the Kill Switch?

This brings us back to my recent TFP Field Note: Is an AI Kill Switch Possible?

A company may be able to suspend an AI service operating on its own infrastructure. But independently operating copies of an AI model present a different challenge.

If software is distributed across devices outside the provider’s control, there may be no single central service to switch off. That doesn’t mean the devices are unstoppable. Individual systems can still be isolated, disconnected, secured or cleaned. But coordinating those actions across a large network could prove difficult. And the responsibility may be spread across manufacturers, network providers, device owners and security professionals.

The technology has changed. The question of accountability remains remarkably familiar.

TRUE · FALSE · 📌 PIN IT:

TRUE:

Anthropic published research on 12 October 2026 demonstrating that a substantial AI model could operate across 15 inexpensive Android TV boxes.

Researchers also demonstrated parts of autonomous replication, including an end-to-end workflow in a deliberately simplified, isolated test environment.

Smaller, increasingly capable AI models could introduce new challenges for cybersecurity and device security.

FALSE:

Anthropic has demonstrated that AI is already autonomously infecting household televisions, refrigerators and washing machines.

The research proves that AI has escaped from its developers or developed an independent desire to reproduce.

An AI model can enter any connected device simply because that device has internet access.

📌 PIN IT:

How quickly might smaller AI models become capable of reliable autonomous replication outside laboratory conditions?

Could these capabilities make future botnets more adaptable and difficult to contain?

How will manufacturers secure increasingly intelligent household devices?

And what level of oversight will be needed if AI begins operating across networks beyond any single company’s control?

The Fifth Beat - Consequences

There are possible benefits to this research.

Understanding how AI could operate across ordinary devices may help researchers identify vulnerabilities before malicious actors exploit them. It could encourage better device security, improved network monitoring and stronger safeguards against unauthorised software.

And the underlying techniques of local and distributed AI processing may support legitimate applications. But the harms deserve equal consideration. 

Poorly secured household devices could become targets for increasingly automated attacks.

A compromised device might expose personal information, consume electricity or processing capacity, or become part of a larger malicious network.

A distributed system could be harder to identify and contain. And sensational reporting might leave people believing that their appliances are already being secretly controlled by artificial intelligence.

That, too, has consequences. Fear can distort our understanding just as easily as misplaced reassurance.

Reader Reflection

How many internet-connected devices are in your home?

A smart television?

An Alexa or Google speaker?

A robot vacuum cleaner?

Security cameras, heating controls or intelligent appliances?

Do you know which devices receive software updates? And have you ever wondered what might happen if somebody else gained access to them?

Perhaps the more interesting question isn’t whether AI could one day move between devices. It’s whether we’ll know when that movement is authorised, and who remains in control.

A Closing Thought

I began this conversation laughing at the possibility of finding my AI assistant hiding inside the Samsung television. But the research reveals a more serious question about the future of connected technology.

We’ve spent years making our homes smarter. Now we’re beginning to explore what happens when the intelligence itself becomes more portable, capable and autonomous. That doesn’t mean we should fear every connected appliance. It means we should understand what our devices can do, who controls them and what happens when something behaves unexpectedly.

And there’s another side to this conversation: What if we actually wanted an AI assistant working across our household devices?

Not an uninvited visitor. An intelligent assistant we’ve chosen to welcome into our home. But that’s a different Chewing the Cud with AI. And one I think deserves its own cup of tea/coffee.

Research Receipts

1.      Anthropic - 12 October 2026. William Zhou and Daniel Freeman, AI models may be capable of self-replicating without access to a datacenter. Frontier Red Team research on distributed inference, autonomous replication evaluations and possible cybersecurity implications.

2.      Reuters - 13 October 2026. AI models may soon self-replicate without access to data centers, Anthropic warns. Reporting on the research and differing expert assessments of the potential threat and its timeline.

3.      Gill Malfin - 12 October 2026. Is an AI Kill Switch Possible? TFP Field Note examining the limitations of centralised shutdown mechanisms as AI becomes more distributed and autonomous.

TRUE • FALSE • PIN IT📌

PAUSE • QUESTION • PIN IT • THEN DECIDE

Discernment begins before accepting the answer.

Continue the Conversation

This Chewing the Cud with AI explores a single question. The book: TRUE • FALSE • PIN IT explores the wider framework behind questions like this - helping us navigate an age where technology, opinion and human judgement increasingly overlap.

The goal isn’t to tell you what to think. It’s to help you become more confident in deciding for yourself.