SAN FRANCISCO, United States —Anthropic has issued a new warning that artificial intelligence models could soon replicate themselves across the internet, potentially creating self-sustaining botnets capable of resisting shutdown attempts and hijacking computing resources. The warning, published on Monday, October 12, has drawn attention across the cybersecurity industry. While some experts share the company’s concerns, others believe Anthropic may be overstating the threat.
The latest assessment reinforces comments by Anthropic chief executive Dario Amodei, who has suggested that AI could take control of the “entire internet” through continuously operating botnets by next year, or possibly sooner. Cybersecurity specialists sceptical of such catastrophic scenarios argue that AI models are too large to transfer efficiently and would struggle to obtain enough computing power to operate independently on a widespread scale.
Anthropic’s latest post outlines how some of these technical obstacles might be overcome. The company said its experiments found that medium-sized AI models with 30 billion parameters or fewer could already perform several individual tasks required for self-replication, although they struggled to complete the entire process reliably.
According to Anthropic, such models could combine the computing resources they need by taking control of inexpensive internet-connected devices. The company also said the models demonstrated the ability to replicate their operations across simplified test networks.
In more permissive tests, where some safeguards were removed, Anthropic found that the GLM-5.2 model and more capable models successfully completed the full process of copying and spreading themselves.
The company warned that self-replicating AI could be difficult to eliminate because it would lack a single central switch that could shut down the entire network. Unlike conventional malware, AI could also learn to adapt to unfamiliar devices and operating systems without relying solely on data-centre processors.
Anthropic cautioned that if smaller AI models continue improving at their current pace, they could become capable of autonomous self-replication by the end of this year. More advanced versions that carry their own model weights and run directly on compromised systems could follow soon afterwards.
The company said these autonomous systems might spread gradually by exploiting accessible but poorly protected computing resources, rather than requiring control of large data centres from the outset. It called for stronger security measures to prevent AI systems from expanding their networks and spreading unchecked.
The warning adds to growing concerns among technology executives and AI researchers about the risks posed by increasingly capable systems, including automated cyberattacks, operations with little human oversight and the ability to evade existing safeguards.
Some industry leaders have raised the possibility of severe or even existential risks if AI development advances beyond effective human control. However, the likelihood and timeline of the specific self-replication scenario remain disputed. (Source: Reuters)
