Your Toaster Is Not Plotting Against You

On October 12, Anthropic published a research post with a title built to make you look sideways at your kitchen: “AI models may be capable of self-replicating without access to a datacenter.” The short version: their researchers think small AI models may be months away from copying themselves onto cheap, poorly secured gadgets like smart ovens and off-brand Android TV boxes, spreading from one device to the next without a human pushing the button.

Here is what they actually did, in plain English. They bought 15 cheap TV boxes, about 55 dollars each, wired them together on purpose, and split a small Google model across all of them. It ran. At 0.39 tokens per second. That’s roughly one word every few seconds. Your grandmother typing with one finger on a flip phone could take this thing in a fair fight.

Then they built a fenced-off test network of eight emulated Android machines with the digital front door left wide open. No password. No lock. Nothing. One open model from a Chinese lab, GLM-5.2, managed to hop from machine to machine and get a smaller model running on each one. But the model doing the driving was too big to live on those little machines, so it ran on a separate server the whole time. And by their own account, the hacking step was made trivial, which they admit makes the result an upper bound. Their bottom line, in their own words: “We don’t yet have a full rogue replicator.” Their forecast is that a small model could reach that level by the end of this year, and they call that a rough estimate.

So, to recap. There is no AI worm in the wild. There’s a lab demo with the doors propped open, a chatbot running at the speed of continental drift, and a prediction. And the headline that came out the other end of the media machine was, more or less, your refrigerator may be plotting against you.

That’s the part a lot of people noticed. Read the coverage and you get smart fridges, smart ovens, vacuum cleaners and lawnmowers listed as the places a rogue AI might hide. Nobody has to say “be afraid of your toaster.” They just have to put the toaster in the sentence. Some folks call that a psyop. I’d call it marketing that knows exactly which nerve it’s pressing. Either way, the effect is the same. You walk away a little more scared of the machines in your own house, and a little more grateful that somebody responsible is watching over them.

Now look at who the bad guys are in this study. The models they tested as would-be worms are open models, the kind anybody can download and run for free. Google’s Gemma. Alibaba’s Qwen. Meta’s Llama. GLM from China. For years the fear talk aimed at open models in general. Now the spotlight has moved to small models, the kind that run on your laptop or your phone without paying anybody a monthly fee. Some people think that’s no accident. Their theory goes like this: if the free, small models that run on your own hardware get painted as tomorrow’s malware, then the rules that follow will land hardest on everyone who isn’t a giant lab with a giant data center. They call it regulatory capture. Pulling up the ladder after you’ve climbed it. I can’t prove intent, and I won’t pretend to. But I’ve watched this pattern long enough to notice it.

And this isn’t the first time the AI worm has come up. The idea has been floating around research papers for a while now, an AI that spreads itself from computer to computer. It makes a great headline. It makes a lousy reality, at least so far. Even outside experts quoted by Dark Reading pointed out that shoving gigabytes of model weights through the weak little connections on these devices is, in their words, “a massive practical hurdle.” Scientific American noted the study stopped short of showing the full threat. Ars Technica called the finding necessarily tempered. The monster under the bed keeps getting described in more and more detail. It just never shows up for the photo.

Blog and social media content by Twin Rivers Communications

Here’s where my own opinion comes in, and I want to be clear that this is commentary. A company that warns us all day long about dangerous AI escaping into the wild might want to look at its own back door first. In late March, Fortune reported that Anthropic had left close to 3,000 internal files publicly reachable, including a draft blog post about an unreleased model called Mythos, the one they say is so good at hacking it’s too dangerous to release. Days later, on March 31, Anthropic shipped an update to its Claude Code tool that accidentally included a map file exposing about half a million lines of its own source code. The company called it human error in packaging, not a breach, and said no customer data or credentials were exposed. It was copied across GitHub within hours anyway. Then in April, Bloomberg reported that a small group in a private online forum got into Mythos itself through a third-party vendor’s environment. Anthropic said it was investigating and had no evidence its own systems were affected.

In my book, that’s a pattern of things walking out the door. And I’d argue it’s a little rich to lecture the public about your smart fridge when your own code went out with the trash. The people who write glowing support for this two-track message, be very afraid of everybody else’s AI, but trust us with ours, deserve some hard questions too.

Now some high-end pondering, the long view. Every powerful technology gets two stories told about it. One is about what it can do. The other is about who should be allowed to have it. The printing press got that treatment. Encryption got it. The early internet got it. And every time, the scariest version of the story was told by the people who stood to gain the most from controlling it. That doesn’t make the danger fake. Cheap gadgets really are poorly secured, and that’s a real problem worth fixing. Anthropic’s own suggested fixes are actually sensible: better default passwords, internet providers watching for strange traffic, manufacturers shipping safer products. Notice that none of that requires banning your laptop model.

So here’s my advice. Change the default password on your TV box. Update your router. And the next time somebody tells you the toaster might be thinking, ask who’s selling the fire extinguisher.

The machines in your kitchen are not plotting against you. At 0.39 tokens per second, they couldn’t plot their way out of a paper bag. And the people who write about this kind of fear-mongering are moving fast.

Sources:

My Living AI News Now. Your Place for AI News. Commentary, not financial advice.

C. Rich

Math Got Solved. Mathematicians Want a Refund.

Nice or Nasty: Which One Makes AI Smarter?

Somebody Just Pulled $30 Billion Off the Table