Oct 12 (Reuters) - Anthropic said on Monday that small AI models could be months away from self-replicating across internet-connected devices, raising the possibility of AI botnets spreading quickly, fighting takedown attempts and even breaking into data centers to turbocharge their growth.
Last month, Anthropic CEO Dario Amodei first floated the idea that within the next year or even sooner, AI might be able to hijack "the entire internet" with "a persistent botnet," a warning that raised eyebrows
in the cybersecurity community.
Botnets -- vast networks of devices controlled by malicious operators -- have been around since the early days of the internet, and are still around.
The network typically includes poorly secured Internet of Things (IoT) devices, such as Android TV boxes, smart appliances and security cameras, and is used by hackers to send spam, seed malware, or sabotage target companies using floods of bogus traffic.
Anthropic's Monday blog post appeared to be the first attempt to articulate how that might happen.
The company said that its experiments showed that a malicious AI agent could stitch together the computing power of inexpensive internet-connected devices to run other AI models. It added that the models were able to 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," the company said.
The report adds to broader concerns about the security risks tied to increasingly capable AI systems, including their potential to automate cyberattacks, operate with limited human oversight and circumvent safeguards.
The advances in models have prompted a series of often dire warnings from industry heads and AI company employees about the possibility of severe, even existential, risks of runaway AI development.
Anthropic said its experiments showed that models with 30 billion parameters or fewer can already perform many of the individual tasks required for self-replication, although they struggle to complete the entire process reliably.
In its simplified tests, where some security steps were removed, Anthropic found that GLM-5.2 and stronger models were capable of end-to-end replication.
Anthropic said that if smaller models continue improving at their historical rate, they could reach the capability required for autonomous replication by the end of 2026.
An AI worm running on low-powered hardware could take days to complete parts of the replication process, Anthropic said. It added that such a network could continue growing, slower than some conventional malware, if compromised devices infected enough additional machines.
Anthropic also warned that future systems could target more powerful computing resources, including data-center processors, although its experiments did not demonstrate autonomous replication through commercial data centers.
(Reporting by Raphael Satter in Washington and Harshita Mary Varghese in Bengaluru; Additional reporting by Gargi Gupta in Bengaluru; Editing by Janane Venkatraman and Shinjini Ganguli)
