TL;DR

  • Anthropic’s red team says small AI models may be only months away from being able to copy themselves onto cheap smart devices like off-brand TV boxes and smart ovens. No datacenter needed.

  • It isn’t happening yet, and if it does, it would be slow. The researchers compare it to a slime mold, not a wildfire.

  • The weak point is the same as always: cheap, poorly secured gadgets in our homes.

How worried?: 🟡 3/5 — Meaningful

TOP STORY

Your smart devices could one day host an AI

Yesterday, Anthropic’s Frontier Red Team published research saying that the internet-connected gadgets in our homes are poorly secured, and they’re starting to have enough computing power to be useful to an attacker.

The researchers asked a new question. Could an AI model spread itself the way a computer worm does, copying itself from device to device, without a datacenter? Their answer is “not yet, but possibly soon.”

What they actually showed:

  • Cheap hardware can run real AI. They bought 15 off-brand Android TV boxes, at about $55 each, and linked them together to run a 26-billion-parameter model. It was very slow, under half a word per second, but it worked.

  • Small models can figure out unfamiliar devices. A 9-billion-parameter model got AI running on 8 of the 9 types of hardware they tested, without being told how. Ordinary malware can’t do this. It’s built for one kind of device.

  • One model can already do the whole loop. In a simplified test network, the open model GLM-5.2 found new machines, copied itself over, and got running again. It reached all 8 machines in 4 of 5 trials. That model is too big to fit on a TV box today. Based on how quickly small models have been improving, the researchers estimate a small model could reach that level by the end of this year.

  • It learns as it spreads. When each copy passed notes to the next, setup success rose from 45% in the first generation to between 61% and 84% in later ones.

Why does this matter?

Last week a post with 2.5 million views showed a router screen with a startling number. A family’s Keurig coffee maker had uploaded about 1,008 GB of data in roughly 10 days, while downloading less than 10 GB. The poster unplugged it and bought his parents a new machine. Nobody has explained what that coffee maker was doing, and there’s no evidence AI was involved. However, it does show that we don't a lot about the type of data that our smart devices are consuming and sending out.

An X Post about a a man who discovers that his parents' coffee maker is using 1 TB of data

An X Post about a a man who discovers that his parents' coffee maker is using 1 TB of data

OUR TAKE

This isn’t a rogue superintelligence. It’s an old problem, insecure smart gadgets, getting a new and more adaptable tenant. The defenses are known: ISPs flagging strange traffic, manufacturers dropping default passwords, and owners updating or replacing old devices. The worry is that most of those defenses are optional today.

How worried?: 🟡 3/5 — Meaningful

WHAT IF….?

What if your smart TV box was quietly working a second job?

Picture a $40 streaming stick behind your TV. It still plays Netflix fine, but at 3 a.m. it’s handling one slice of an AI model that’s spread across 14 other strangers’ living rooms. The model is slowly scanning for its next home.

You’d never notice, except perhaps for a router chart that suddenly shows your “appliance” sending out a terabyte of data.

OTHER NEWS

Why is everyone suddenly saying “superintelligence”?

WHAT DO YOU THINK?

Login or Subscribe to participate

Hit reply: are we too worried, or not worried enough? We read every response.

Recommended for you

View all
caret-right