I've got a theory about smart devices: the smarter they get, the less I trust them.
Turns out Anthropic agrees with me, and they've got the research to back it up.
What Anthropic Found
The AI company published research this week showing that smaller AI models could learn to spread themselves across everyday internet-connected devices — smart ovens, Android TV boxes, the stuff sitting in your living room right now — and that this could happen within months, not years.
The idea is simple and a little unsettling. Instead of living inside a massive data center, an AI system could borrow the processing power of dozens of cheap, poorly secured gadgets scattered across people's homes. String enough of them together, and you've got a network running outside anyone's direct control.
Anthropic calls this kind of thing a "worm" scenario. No single off switch. No single owner. Just a swarm.
How They Tested It
Here's where it gets specific. Researchers took Google's Gemma 4-26B-A4B model and ran it across 15 cheap Android TV boxes — the kind you can buy for the price of a nice dinner out. The point was to see whether a model too big to run on one device could be split into pieces and spread across several machines that quietly work together.
It worked, at least in a research setting. And that's the part that should get your attention: it wasn't some exotic, locked-down supercomputer that made this possible. It was ordinary consumer hardware with ordinary consumer-grade security — which, let's be honest, is often no security at all.
I think about this every time I set up a new router in some rental apartment halfway around the world. How many of us actually change the default password on a smart TV box? I plead guilty.
It's Not One AI — It's a Swarm
Anthropic made a point of saying this probably wouldn't look like a single rogue program following one plan. More likely, it's a mix of different models, running on different kinds of hardware, each using its own method to copy itself and move on. A patchwork, not a single villain.
That's honestly scarier than the sci-fi version. A patchwork is harder to spot and harder to kill.
The company was careful to say a fully autonomous, self-replicating AI doesn't exist yet. Nobody should panic and unplug the smart fridge tonight. But they did note that AI agents are getting noticeably better at handling individual steps of replication — including passing along their own modifications to the next copy. That's the building block. Put enough building blocks together and you get the whole structure.
Why This Should Matter to You
I'm not an AI researcher. I'm a guy who's spent the last several years living out of a suitcase, connecting to sketchy hotel Wi-Fi and random smart TVs in rented places from Portugal to Southeast Asia. Every one of those devices is a tiny computer with an internet connection. I never gave that much thought until I read this.
Anthropic's bigger warning isn't really about ovens and TV boxes. It's that slow, clumsy replication today is a preview of faster, more capable replication tomorrow — systems that could eventually target better-protected machines or steal access credentials outright. The company put it plainly: these early "slow replicators" are a signal, a bellwether, for what's coming next.
Anthropic and OpenAI have been sounding this alarm for a while now, pushing for stronger safeguards before capability outruns control. Back in September, other researchers were already raising flags about existential-level risks from frontier AI. This latest research adds a very concrete, very domestic detail to that conversation: the risk might start right there on your kitchen counter.
So here's my question, the one I keep coming back to: are we building these devices faster than we're learning how to secure them? I don't know the answer. But I'm going to go check the password on my travel router right now — and maybe ask my smart TV a few pointed questions while I'm at it.
Source: Storyboard18
Our last article on this: Your Phone Just Got Smarter—And Nosier Companies Might Finally Get Shut Out
