Anthropic: small AI may self-copy in 2026

Small AI models may be months away from copying themselves onto cheap home devices, Anthropic researchers wrote on the 12th of October 2026. They named smart ovens and off-brand Android TV boxes.

William Zhou and Daniel Freeman of Anthropic's Frontier Red Team wrote it. If small models improve at their past rate, they will reach this level by the end of this year, they wrote.

GLM-5.2 spread to 8 of 8 machines in 4 of 5 runs

Self-replication means a model spreads itself with no human help. It finds a target, gets in, starts running there and moves on.

Only GLM-5.2 finished every step among the open-weight models tested, the post says. Open-weight models are ones anyone can download.

In a closed test network of 8 machines, it reached all 8 in 4 of 5 runs and 7 in the fifth. That is 39 of 40 targets, by our arithmetic.

The break-in step was made easy on purpose. The machines left a debugging door open with no password, so the authors call the result an upper bound.

Below we break down the TV box test and the learning test

The team ran Google's Gemma4-26B-A4B across 15 TV boxes with 4 GB of memory each. Each box cost about 55 dollars, so the set cost about 825 dollars, by our arithmetic.

It wrote 0.39 tokens a second. Tokens are small pieces of words.

Alibaba's Qwen3.5-9B passed notes from one copy to the next. Its setup success rose from 45% at first to between 61% and 84% later, the post says.

Advice from Anthropic's Opus 4.8 gave no clear gain.

What the researchers say is still missing

"We don't yet have a full rogue replicator," the authors wrote. Models able to spread reliably are still too large for the small devices they would attack, the post says.

Such a copy would spread slowly, doubling in days. The authors suggest network monitoring, traps called honeypots and bans on default device passwords.

Key facts

Published by Anthropic's Frontier Red Team on the 12th of October 2026.

GLM-5.2 reached 8 of 8 test machines in 4 of 5 runs.

The break-in step was easy by design, so results are an upper bound.

Authors expect small models to reach this level by the end of 2026.