Anthropic Warns Small AI Models Could Self-Replicate on Home Appliances Like Computer Viruses

Anthropic has warned that even small artificial intelligence (AI) models could reach a level as early as the end of this year where they operate like "computer viruses," copying themselves and spreading across household appliances.

On the same day, an Anthropic co-founder advanced the projected timeline by a year, stating that AI reaching the stage of self-improvement could become possible next year.

In a research report released on October 12 (local time), Anthropic stated, "It may be only a few months before small AI models acquire the capability to self-replicate across Internet of Things (IoT) devices, such as smart ovens or low-cost Android TV boxes."

Anthropic explained that because such self-replicating AI spreads and operates scattered across numerous devices, it is difficult to stop because there is no central shutdown point—unlike general AI services where turning off a few servers is enough.

Calling this an "AI worm," Anthropic conducted a three-stage verification experiment.

First, they purchased 15 low-cost TV boxes priced at about $55 (approx. 75,000 KRW) each, connected them, and installed Google's AI model "Gemma 4" across them.

Although the speed was slow, the AI operated.

Second, they tested whether the AI could independently install and operate another AI model on an unfamiliar device.

Even a small AI model with 9 billion parameters succeeded on 8 out of 9 tested devices.

Finally, they created a test network using 8 virtual devices and had the AI planted in the first device autonomously repeat the process of locating, installing, and operating AI on the other devices.

In 5 rounds of testing, only Zhipu AI's "GLM-5.2" from China succeeded in migrating to all 8 devices in 4 rounds, and up to 7 devices in 1 round.

However, because GLM-5.2 was large in size, it received instructions from an external server, and the device security had been intentionally set to be lax.

While acknowledging that a fully self-replicating AI does not yet exist and that the speed of replication spread would be slow, Anthropic projected that under current development trends, small models could reach the level of GLM-5.2 by the end of this year and become capable of direct self-replication inside devices.


Anthropic explained that a realistic worst-case scenario is a "swarm" form where multiple types of AI mix various devices and replication methods.

Going beyond the scope of the experiments, the company stated it anticipates "a future where agents run on all kinds of hardware, including things outside of human control."

Jack Clark, co-founder and head of policy at Anthropic, stated in a newsletter titled "Import AI" published the same day, "It is more likely that full recursive self-improvement (RSI) will become possible next year," adding, "Previously, I thought 2028 was more likely."

Clark based his assessment on test results from the AI safety research organization P-Zero Research, noting that for AI to improve itself, a researcher's "intuition" in choosing which experiments to run and interpreting the results is key, and this capability is improving rapidly.

In these tests, AI models designed experiments and interpreted results for actual AI development tasks such as training data selection and language model training. This capability has doubled approximately every three months since December of last year.

Anthropic's "Opus 5.5," which achieved the highest score, surpassed the scores of human experts who had worked at OpenAI, Google DeepMind, and elsewhere.

However, P-Zero researchers stated that because the tasks were easy and of a type that could be graded quickly, making it easy for AI developers to prepare custom countermeasures, the rate of improvement may have been exaggerated.

Expert reactions were mixed.

According to Reuters, Ciaran Martin, former head of the UK National Cyber Security Centre (NCSC), evaluated the Anthropic report, saying, "Succeeding in running a small model on a cheap TV does not constitute a security crisis," and calling it "an interesting technical paper mixed with no small amount of vendor hype."

Jessica Ji, a senior research analyst at Georgetown's Center for Security and Emerging Technology (CSET), said, "We don't know yet how likely it is that such a thing will actually happen," adding, "It is not zero, but it is not certain either."

(Photo: Yonhap News, Getty Images)
※ Please note: This article was translated by AI and may contain errors.