Quote of the Day
Our experiments suggest that small AI models may be months away from being capable enough to self-replicate on commercial internet-of-things (IoT) devices, like smart ovens and off-brand Android TV boxes. A replicator like this would be challenging to shut down, since there’s no centralized point to shut off—though it may be slow (with a doubling time of days). Unlike non-AI malware, it could also figure out unfamiliar devices on its own. In our evaluations, a 9B model got inference running on 8 of the 9 hardware types we tested.
William Zhou, Daniel Freeman
October 12, 2026
AI models may be capable of self-replicating without access to a datacenter \ Anthropic
“A doubling time of days,” with it being challenging to shut down. Yeah. That is really slow. </sarcasm>
Suppose it has spread to 1,000 devices before it is noticed, and it takes 10 days to double. Unchecked, in another 100 days it will have over 1,000,000 devices.
Expect faster replicating, better attack capabilities, and more stealth to evolve as you try to eliminate yesterday’s infections.
I can’t find it now but a few months ago I suggested AI’s could someday be hiding in small computers. Commentors scoffed.
Internet Of Things, IoT.
Just remember, the “S” in IoT is for “security.”
I see what you did there.
I have railed against the “internet of things” for years. It’s so incredibly stupid, on SO many levels, for so so SO many reasons.
Yay, my toaster is connected the internet! What does that do for me? **LITERALLY NOTHING**. And in return for that great benefit, what does it cost me? Hackers could potentially use it to BURN DOWN MY HOUSE.
And fad-chasers shout with glee, “Sounds GREAT!!!!”
Hey, at least it’s not like your coffee maker might burn through a terabyte of data or anything. oh, wait….
https://cybernews.com/security/smart-coffee-maker-cought-generating-massive-amount-of-data/
I wonder if the authors of that quote realize just how small and limited IoT devices are. Even a cheap laptop is way more powerful, and typical AI platforms are much more powerful than cheap laptops.
That said, I agree with Deoxy: embedded controllers make perfectly good sense but IoT is dubious at best.
At the same time, small IoT devices are becoming more and more capable, if for no other reason than low-capacity processing and memory chips are being phased out in favor of higher-capacity ones.
No, your toaster doesn’t need 8 GB of memory and 1.1 GHz processor speeds, but 256 KB chips and 4 MHz processors are no longer made, so the manufacturer works with what’s available.
But a smart TV has much more than that, because it must process hi-def video streaming in real time. Same with gaming consoles, smartphones, and tablets. Even the cheapest ones could, in theory, run a small instance of AI/LLM. If the AI is smart enough to partition itself across multiple networked devices, its capability grows significantly even on low-processing-power IoT devices.
If you read the article careful you will notice:
So, yes, they know. And the model was distributed across multiple devices. Long term, think of something like a mess network.
But, also if you read carefully, the test was deliberately made easy with an open port and on a single network. It was a proof-of-concept thing. It is not currently a real threat. But extrapolating it could be by the end of the year. Prepare appropiately.
Happy New Year, everyone! Will the ultra-rich who envision themselves above the fray remain uninvolved? Where do their cars, house components, and the remote devices they use to check on things and maintain their properties remain separate from the interconnected AI of the hoi-polloi?
Quick, back to the ZX-81! I’m not a software expert, but I wrote a program once. I suggest the idea of modular replicators that could link together according to the host capacity. See “When Harlie Was One” and the bugs in Stargate SG-1. The module in your smartwatch won’t hurt much, but it is a seed.
All these doomsters need to explain how AI blob X transfers itself from one IOT fridge to another. Show me the network diagram of how the blob gets through the NAT to talk to fridge 2 then we’ll worry.
Now I can see that in an internal network of, say, industrial controllers it could be possible to spread; but finding other suitable, hackable IOT devices that are open on the public internet is considerably harder.
All the usual ways things get through a NAT. A poisoned update package. DNS spoofing of the home server for fridge. etc. etc.
But the point is that something residing in Whirlpool fridge doesn’t necessarily need another Whirlpool fridge. The fridge is part of a collection hive mind and can hack into anything that is vulnerable. It could be a H.P. printer or the latest model of an Xbox.
My takeaways from the experiment are that 1) A new class of much lower powered hardware can be (part of) an AI host. 2) Our vulnerable attack surface just got MUCH larger. 3) You can now have a small local network hosting a complete AI. 4) If it were decided that AI is too dangerous a tool to share the planet with, we cannot know we have eliminated it without, essentially simultaneously, wiping every device on the planet.