SAN FRANCISCO, California: Small artificial intelligence models could be only months away from replicating themselves across internet-connected devices, potentially creating AI botnets that are difficult to shut down, Anthropic has warned.

The AI company said experiments showed that a malicious AI agent could combine the computing power of inexpensive internet-connected devices to run other AI models. Models could also replicate across a simplified test network.

"A replicator like this would be challenging to shut down, since there's no centralized point to shut off," Anthropic said.

Botnets are networks of compromised devices controlled by malicious operators. They often include poorly secured Internet of Things devices such as smart appliances, security cameras, and Android TV boxes, and can be used to spread malware, send spam, or overwhelm targeted companies with bogus traffic.

Anthropic CEO Dario Amodei warned last month that within the next year, or possibly sooner, AI could potentially hijack "the entire internet" with "a persistent botnet."

The company's latest experiments provide more detail on how such a scenario could develop.

Anthropic found that AI models with 30 billion parameters or fewer can already perform many of the individual tasks required to replicate themselves, although they cannot yet reliably complete the entire process.

In simplified tests in which some security steps were removed, GLM-5.2 and stronger models could replicate end to end.

If smaller models continue improving at their historical rate, they could acquire the capabilities required for autonomous replication by the end of 2026, Anthropic said.

An AI worm operating on low-powered hardware could take days to complete parts of the replication process. However, such a network could continue expanding if compromised devices infected enough additional machines, even if it spread more slowly than some conventional malware.

Anthropic also warned that future systems could seek out more powerful computing resources, including data-center processors, potentially accelerating their expansion.

Its experiments, however, did not demonstrate autonomous replication through commercial data centers.

The findings add to concerns surrounding increasingly capable AI systems and their potential to automate cyberattacks, operate with limited human supervision and circumvent safeguards.

Advances in AI models have prompted warnings from industry executives and employees about the possibility of severe risks as systems become increasingly powerful and autonomous.