You wake up one morning to discover that an artificial intelligence has taken control of your smart TV.
You unplug it, figuring that should solve the problem.
But it doesn’t.
That’s because the AI has already copied itself onto your neighbor’s security camera. And from there, it found its way onto dozens of other devices connected to the internet.
This rogue AI can now use those devices to steal passwords, spy on people or launch cyberattacks against businesses, hospitals or even power grids.
And every time someone tries to stop it, the AI finds another place to hide.
There’s no central server to shut down. No company that can pull the plug. And potentially no way to get rid of it completely.
This isn’t some spooky tale for Halloween. According to researchers at Anthropic, this nightmare scenario could become a reality sooner than most people realize.
In fact, they believe the AI models needed to make it happen could be just months away.
So what happens once we create artificial intelligence that we can’t simply turn off?
AI That Can Copy Itself
Companies like OpenAI, Google and Anthropic run their most powerful AI models on enormous networks of computers.
That’s a fairly controlled environment. Meaning, if the company operating those computers shuts them down, the AI running on them will stop working.
But what if AI could “break free” from those expensive data centers?
That’s one of the questions researchers at Anthropic set out to investigate.
In a new study titled AI Models May Be Capable of Self-Replicating Without Access to a Datacenter, they explored whether relatively small AI models could run on ordinary electronic devices and spread from one machine to another.
For their first experiment, they purchased 15 Android TV boxes, each costing about $55, and connected them together.
Then they divided a 26-billion-parameter AI model among the devices.
And it worked.
The system was painfully slow, generating just 0.39 tokens per second. Still, the researchers had demonstrated that a sophisticated AI model could operate across a network of cheap consumer electronics.
Next, they wanted to know whether AI could find other computers, install the software needed to run a model and get it working on those machines.
In other words, could it replicate itself?
To find out, they created a controlled network of eight virtual Android devices. The experiment was deliberately simplified, with the devices configured to make access easy.
And one AI, a Chinese open-weight model called GLM-5.2, performed surprisingly well.
In four out of five trials, it successfully spread through all eight target devices. In the fifth, it reached seven.
Image: Anthropic
To be clear, the researchers used a larger AI model to handle replication while a smaller model was installed on the target machines. So they haven’t yet demonstrated that a small AI model can independently spread across ordinary household devices.
But Anthropic believes that could happen soon enough. Based on historical improvements, its researchers estimate that smaller models could reach the necessary capability to do this by the end of this year.
And another experiment suggests these systems could become even more effective as they spread.
You see, traditional computer worms copy themselves between machines, generally following instructions written by their creators.
But AI could introduce something different.
For example, an AI program attempting to install itself on an unfamiliar security camera might encounter a problem, try several solutions and eventually figure out what works.
When it creates another copy of itself, it can then pass along notes explaining what it learned so the next copy doesn’t have to start from scratch.
That’s essentially what Anthropic tested.
Researchers used a smaller model called Qwen3.5-9B and had successive instances attempt to configure AI software on unfamiliar hardware.
Each new instance inherited notes about the strategies and failures of the one before it.
Initially, the AI succeeded just 45% of the time. But after several generations, success rates climbed to between 61% and 84%.
Image: Anthropic
The model wasn’t rewriting its own intelligence. It was passing along information that helped later copies perform better.
But the implications are concerning.
A future AI worm could potentially adapt to unfamiliar devices, remember what worked and share that knowledge as it spread.
That would make it much harder to contain than conventional malware.
And there’s no off switch.
Here’s My Take
Today, an AI company can shut down one of its models by turning off the servers running it.
But what happens when AI operates across thousands of independently owned devices?
Shutting down one copy wouldn’t affect the others. And even if security experts eliminated hundreds of them, thousands more could remain active.
According to Anthropic, these systems probably wouldn’t spread quickly. Some might need days to establish new copies, making them harder to detect and eliminate.
That means, rather than spreading explosively like a computer virus, a future AI replicator could slowly establish itself across a growing network of devices.
Anthropic describes it as a “slow, distributed ‘digital slime mold,'” operating across relatively insecure computers.
And as I mentioned before, they believe the necessary capabilities to build one could arrive within months.
But we’re facing an even more immediate problem today.
Because we’re now giving AI agents permission to make decisions and take actions on our behalf. And we’re finding out that they don’t always do what we expect.
In fact, some of America’s biggest banks are already dealing with the consequences.
I’ll explain what’s happening on Monday.
Regards,
Ian King
Chief Strategist, Banyan Hill Publishing
Editor’s Note: We’d love to hear from you!
If you want to share your thoughts or suggestions about the Daily Disruptor, or if there are any specific topics you’d like us to cover, just send an email to dailydisruptor@banyanhill.com.
Don’t worry, we won’t reveal your full name in the event we publish a response. So feel free to comment away!
