New to this? Read it in simple words
- Anthropic says small AI models may soon be able to copy themselves onto cheap home devices.
- In a test, one AI model ran across 15 TV boxes that cost about $55 each.
- Only one large open model did the whole copying chain, and only in a simplified test.
- No AI worm like this has been seen in the wild yet.
- Worm
- A harmful program that copies itself from one computer to the next without anyone’s help.
- Botnet
- Many hacked devices that one attacker controls together, often without their owners knowing.
- Open model
- An AI model whose files anyone can download and run on their own machines.
A computer worm copies itself from one machine to the next without human help. Anthropic’s Frontier Red Team asked whether an AI model could do the same, carrying its own files with it.
What Anthropic tested
Anthropic published the research on October 12. Its team ran Google’s open model Gemma4-26B-A4B split across 15 off-brand Android TV boxes, bought for about $55 each.
The model wrote 0.39 tokens per second, which is very slow, but it ran. These boxes matter because a botnet called Kimwolf once took over more than 2 million devices, mostly such boxes.
In another test, a 9-billion-parameter model, Qwen3.5-9B, got AI software running on 8 of 9 kinds of hardware it was given.
The steps Anthropic tested on October 12, in a simplified network where the hacking step was made easy.
The whole chain, in a simplified test
The team then built a closed network of 8 cloud machines running an Android emulator. The worm moved through a debugging port that needed no password, so the hacking step was made easy.
Only GLM-5.2, a much larger open model, completed the whole chain. It reached all 8 machines in 4 of 5 trials, and 7 machines in the fifth.
Smaller models could do many single steps, but made mistakes when they had to chain them. Anthropic calls the test “an upper bound on real-world success”.
How soon, and what could help
“If small models improve at their historical rate, they’ll match this capability level by the end of this year,” the researchers wrote. They call this timeline only a rough estimate.
Such a worm might spread slowly, with its numbers doubling over days, Anthropic says. It names smart ovens and fridges as possible hosts, but it did not test them.
Anthropic suggests that internet providers watch for unusual traffic and cut off infected customers, and that device makers ship safer default settings. No AI worm like this has been seen in the wild so far.
Sources
Every fact in this story comes from the sources below. Open them to check our work.
- 1Primary source · October 12, 2026AI models may be capable of self-replicating without access to a datacenter Anthropic
- 2Research · October 13, 2026Anthropic warns AI models could soon copy themselves across home devices CNBC
- 3Research · October 13, 2026Anthropic says small AI models may soon copy themselves onto smart devices TNW
- 4Research · October 13, 2026Anthropic warns of self-replicating AI worm threat that doesn’t need data centers to self-improve and spread Tom’s Hardware
The tests and the timeline are Anthropic’s own, and the company calls the timeline a rough estimate. No outside group has repeated the tests yet.
Anthropic’s tests suggest small AI models could soon spread like worms across cheap home devices, though none has been seen in the wild. How we rate the mood