Anthropic says small AI models may soon copy themselves

Small AI models may be months away from copying themselves onto cheap devices like smart ovens and Android TV boxes, Anthropic’s Frontier Red Team says. No such replicator exists yet, it says.

Small AI models may be months away from being able to copy themselves onto cheap internet-connected devices, Anthropic said in an Oct. 12 research post.

The targets would be devices like smart ovens and off-brand Android TV boxes, the post says. No data center needed.

What Anthropic found

The company’s Frontier Red Team reported these findings:

Cheap hardware: It ran a 26B model across 15 off-brand Android TV boxes, about $55 each, at 0.39 tokens per second.
Setup: In its evals, a 9B model got inference running on 8 of the 9 hardware types tested.
Full cycle: In simplified evals, GLM-5.2 and stronger models are capable of end-to-end replication, the post says. Models of 30B parameters and under can do many single steps but make mistakes chaining them.
Learning: Small models passing notes to later copies got better at setup, from 45% success to between 61% and 84%.

If small models improve at their historical rate, they’ll match that capability by the end of this year, Anthropic says. It calls that a rough estimate.

Why it matters

A replicator like this would be hard to shut down, the post says. There is no central point to shut off. It may also be slow, with a doubling time of days.

Unlike non-AI malware, it could figure out unfamiliar devices on its own, Anthropic says.

The tests kept hacking steps trivial, which Anthropic calls an upper bound on real-world success. It says it does not yet have a full rogue replicator.

Who it binds

No one. This is research, not a rule. Anthropic points to regulation as one lever: default IoT passwords are now non-compliant in the U.K.

What it means for you

Parents and small business owners: change default passwords on smart TVs, cameras and appliances. Anthropic notes past botnets spread by guessing manufacturer default logins.

Founders building IoT devices: ship secure defaults. Anthropic says manufacturers can shrink the pool of targets.

What’s next

Anthropic says the spread would leave a network signature internet providers could spot. It points to quarantining infected customers: one ISP’s quarantine cleared 92% of Mirai infections within 14 days, compared with 74% with no notification.