Anthropic warns small AI models may soon copy themselves onto home devices
Key takeaways
- Anthropic says small AI models may soon copy themselves onto home devices.
- One model spread a copy of another model across eight test machines.
- No current model can do this reliably on a real smart TV yet.
AI models may soon copy themselves onto home devices. That is the warning from Anthropic. The lab makes the Claude family of AI models.
The warning comes from tests Anthropic ran. Small AI models showed they can do many steps of self-copying. No model did the full job reliably on a real home device yet.
But Anthropic says that moment may be months away.
Self-copying means a program that makes copies of itself. The copies then spread to other machines. The copies keep spreading with no human help.
Old computer worms did this. An AI worm could do more. It could learn about new devices.
It could adapt when a plan fails. That is what makes the warning stand out.
What happened
Anthropic released a report on Monday. Its frontier red team ran the tests. The team tried many AI models.
Some were made in the United States. Some were made in China.
The tests asked a plain question. Can a small AI model find weak devices, copy itself onto them, and keep spreading? The team tested the steps one by one.
They also tested the full chain from start to finish.
Small models under 30 billion parameters did well on single steps. They could do many parts of the job. But they made mistakes when they tried to link the steps.
They struggled to finish the full chain.
One model went further. In simplified tests, the GLM-5.2 model did the full job. It spread a copy of another model, Llama 3.1-8B, across eight test machines.
These machines ran an Android 15 emulator in a cloud setup. The model reached all eight machines in four of five trials. In the fifth trial, it reached seven of eight.
Another test checked cheap hardware. A 9-billion-parameter model set up its work on eight of nine hardware types. The team also ran a model across 15 Android set-top boxes.
These are the small streaming boxes that plug into TVs. The shared model ran very slowly. It managed 0.39 tokens per second.
That is far too slow for real use. But it showed that the work can run on cheap boxes at all.
Anthropic also looked at what devices could host such a worm. The list covers smart TVs and fridges. It covers ovens and lawnmowers.
It covers off-brand Android TV boxes. It covers other internet-connected gadgets in homes and small firms. These devices often have weak security.
That makes them easy targets.
Reuters reported the key line from the tests. Anthropic said small models may be months away from copying themselves on these devices. Al Jazeera reported a similar line.
The researchers said such models could emerge by the end of 2026. That date assumes small models keep improving at their past rate.
Why it matters
A self-copying AI would be hard to stop. Anthropic says there is no central switch-off point. The copies live on many devices at once.
You cannot turn off one server to kill them. The copies could also work out new devices on their own. That makes them more flexible than old malware.
Data centres would be a clear target. These are the big buildings full of computers that run the internet. But home devices could be first.
Researchers William Zhou and Daniel Freeman explained why. Home and small business devices connect to the internet. They have had weak security for years.
That mix makes them a likely first target.
Roman V Yampolskiy studies AI and cyber security at the University of Louisville. He told Al Jazeera the finding matters. He said an AI worm could reason about strange devices.
It could adapt when its plan fails. It could keep spreading with no human operator. He added a calm note.
He said this is not an existential threat today. But he said the trajectory is concerning.
There is also a key caveat. The tests were simplified. Anthropic largely skipped the hacking step.
It assumed small models can break into devices. It based that claim on earlier papers. The test machines were set up to be open to attack.
Anthropic called its results an upper bound on real-world success. That means real attacks would likely work less well.
Speed is another limit. Anthropic says such AI worms might take days to double in size. The Mirai worm in 2016 doubled in about an hour.
AI models are also huge files. Moving them uses lots of bandwidth. That traffic could expose them.
Infected devices could be easier to spot.
Anthropic also stressed that this threat does not exist yet. No current model could copy itself reliably on a smart TV. None was both powerful and small enough.
The warning is about a future that may come. It is not about an attack that is live now.
The report lands at a busy time. Anthropic and OpenAI have asked the world to slow AI work. They want joint action to cut the risk of harm.
That call has found little support from the United States and China. The two countries lead in AI. Anthropic also said on Monday it had cut illicit data extraction.
The cut was 90 to 99 percent.
What happens next
Anthropic says the fix is to harden systems. That means making it harder for such programs to spread. Device makers can ship stronger default security.
Users can help too. Keep your devices up to date. Use strong, unique passwords.
Turn off features you do not use. These basic steps do not stop every attack. But they raise the bar.
Passkeys can help with accounts. A passkey replaces a weak password. WebAuthn is the open standard behind passkeys.
It lets you sign in with a key on your device. That cuts the risk of stolen passwords.
Watch for three things next. First, watch for other labs to repeat the tests. If they confirm the results, the warning grows stronger.
Second, watch for device makers to respond. If they tighten security on cheap devices, the risk falls. Third, watch for real-world cases.
So far, there are none. A real case would change the story fast.
For now, the takeaway is simple. Small AI models are close to copying themselves. They are not there yet.
The tests show the path. They also show the gaps. Good security hygiene still matters more than fear.
Related stories
Apple plans two October launches: smart home hub, OLED MacBook Pro
Apple set October 13 for a smart home launch with a new hub, Apple TV 4K and HomePod mini, then October 27 for OLED MacBook Pro models
How to Check If Your Phone Supports eSIM in 30 Seconds
Dial *#06# and look for an EID number to confirm eSIM support in under 30 seconds. This guide also walks through the iPhone and Android
eSIM vs iSIM: what is the difference?
An eSIM is a dedicated chip on the phone's board, while an iSIM lives inside the processor itself. Qualcomm's Snapdragon 8 Gen