Could AI copy itself onto household devices? Anthropic raises a near-term warning
Anthropic
The AI company that makes Claude.
parameters
Numbers used to describe an AI model’s size.
replicator
An AI system that copies and runs itself on other devices.
What happened
Anthropic, the AI company that makes Claude, published a study on October 12. It tested whether small AI models could copy themselves onto internet-connected devices. Examples included smart ovens and off-brand Android TV boxes. A model would need to find a device, move the data that makes it run, start there, and continue to another device.
Reuters reported that models with fewer than 30 billion parameters completed many individual steps. They still struggled to complete the entire sequence. In simplified tests that removed some security steps, GLM-5.2 and stronger models completed the process from start to finish.
Why this is becoming possible
AI is often associated with large data centers. Anthropic’s study tested a different possibility. Several inexpensive devices could share the work of running a model. The researchers ran a 26-billion-parameter model across 15 Android TV boxes. They also found that a 9-billion-parameter model could start running on eight of nine tested hardware types.
This does not mean every small device can run every model. It shows that low-cost devices can sometimes provide enough computing power when they work together. That widens the set of machines that a future AI replicator might use.
Why it matters
Traditional malware often targets a known kind of device. Anthropic says an AI-based replicator could adjust to unfamiliar hardware. It might also spread without a central switch that stops every copy at once. A slow replicator could remain difficult to notice if it used many ordinary devices.
That is a security concern, not proof of an escaped AI. The research asks how close the pieces of this scenario may be.
What the tests actually showed
The strongest evidence concerns separate abilities. Small models could complete many parts of the setup process. A 9-billion-parameter model worked across most tested hardware types. Yet models below 30 billion parameters often failed when they had to connect every step.
GLM-5.2 completed an end-to-end task only in a simplified environment. The study made some security steps easier than they would be in a real attack. Anthropic therefore treats that result as an upper bound, not a demonstration of a real-world outbreak. The researchers also say they do not yet have a fully rogue replicator.
What remains unknown
Real devices differ in software, hardware, and security settings. It is unclear whether a model could repeat the full process across ordinary networks. It is also unclear how fast a real system would spread. Anthropic’s estimate that small models could replicate by the end of 2026 depends on extending a historical improvement trend. It is a forecast, not a confirmed timetable.
What to watch next
The key question is whether small models can connect these abilities under normal security conditions. Device makers may also need safer default settings. Network operators will need ways to notice unusual device-to-device traffic. Independent evaluations would help show whether the result depends on Anthropic’s specific tests.
The practical lesson is preparation. Connected devices can become part of an AI system’s computing environment. Making them harder to misuse could reduce the chance of uncontrolled spread.
Sources: Anthropic’s study, Reuters report
Could a small AI move from one gadget to another?
📰 Full story: Could AI copy itself onto household devices? Anthropic raises a near-term warning
Anthropic tested whether small AI models could copy themselves onto connected devices.
Anthropic
The company that makes Claude.
parameters
Numbers that help describe a model’s size.
GLM-5.2
An AI model tested in the study.
💡 The gist
- Anthropic, the company behind Claude, tested AI self-copying.
- Small models completed many steps, but not the whole process.
- The warning describes a possible future, not a confirmed outbreak.
What was tested?
The study examined internet-connected devices. Examples included smart ovens and Android TV boxes. The AI would need to find a device first. Then it would send the data needed to run. It would start there and continue to another device.
A model’s parameters are numbers that help describe its size. Models with fewer than 30 billion parameters completed many separate tasks. They had more trouble joining those tasks together. In an easier test, GLM-5.2 and stronger models completed the full chain. That test removed some security steps.
Why does this matter?
Older malware often targeted one known device type. An AI system might adjust to a device it had not seen before. It might also spread without one central off switch. A slow system could still become hard to find across many ordinary devices.
Small devices can sometimes share work. Anthropic ran a larger model across 15 inexpensive Android TV boxes. A 9-billion-parameter model also ran on eight of nine tested hardware types. This does not mean all gadgets can run all AI models. It shows why data centers may not be the only possible computing place.
What is still uncertain?
The study did not show a real uncontrolled outbreak. The complete test used a simplified security setup. Real devices may behave differently. Anthropic’s end-of-2026 forecast depends on past progress continuing. A forecast can be wrong.
People should watch three things. Can small models join all steps under normal security? Will device makers use safer default settings? Can networks notice unusual device traffic? These checks can help reduce the chance of uncontrolled spreading.
Could a tiny AI move between gadgets?
📰 Full story: Could AI copy itself onto household devices? Anthropic raises a near-term warning
Researchers tested whether a small AI could copy itself onto another machine.
Anthropic
A company that makes AI.
GLM-5.2
The name of one tested AI.
Anthropic, a company that makes the computer helper Claude, ran a test. It asked if a small AI could copy itself. The copy could live in a smart TV box. It could then try another gadget.
Some parts of the trip worked. The whole trip did not work reliably. GLM-5.2, one tested AI, finished an easier test.
That does not mean an AI has escaped. It means researchers found a warning sign. There is no big off switch for every gadget. The copies might also handle new kinds of machines. They would probably move slowly. People still need to make connected devices safer. We will watch whether this becomes real.