Anthropic warns that artificial intelligence (AI) could soon self-replicate across cheap connected devices. Its tests showed limited capabilities in small models, but the company has not built a fully autonomous replicator.
Researchers William Zhou and Daniel Freeman published the findings in a company research post on Monday; they work in Anthropic’s Frontier Red Team.
Models with 30 billion parameters or fewer completed many individual replication tasks, according to the company. However, they often failed when combining those tasks into a complete sequence.
In a simplified network experiment, a more capable model completed the replication tasks. Researchers removed key security barriers and ran the controlling model on a separate server.
Anthropic’s team separately ran a model across 15 Android TV boxes that it purchased and connected. That experiment demonstrated how inexpensive devices could collectively supply computing power.
The company estimates that replication across such hardware could be slow, with the number of copies potentially doubling over days.
Its concern centres on persistence. A distributed system could lack a central service that defenders could disable to stop it.
Anthropic suggested that smaller models could reach the necessary capability level by the end of 2026 if historical improvement rates continue. Researchers described that timing as a rough estimate based on general model capabilities.
They also outlined a possible threat to data centres, where compromised processors could provide more computing resources. Their experiments did not establish that such attacks were happening.
Read: Anthropic Says Claude Sent Fake Murder Tip To Philadelphia Police
Anthropic chief executive Dario Amodei had warned in September that AI could create a persistent botnet across the internet within a year or sooner.
Ciaran Martin, former head of Britain’s National Cyber Security Centre, told Reuters the experiments did not establish an immediate security crisis. He saw “more than a hint of vendor exaggeration” in the warning.
Zhengxiong Li, an assistant professor of computer science and engineering at the University of Colorado Denver, expressed greater concern. He compared the scenario to a computer worm that could improvise ways to compromise unfamiliar devices.
Li argued that spreading across numerous machines without a central service could make such a system difficult to contain.
Jessica Ji, a senior research analyst at Georgetown University’s Centre for Security and Emerging Technology, considered autonomous spread a reasonable possibility to examine. But she cautioned that researchers do not yet know how likely it is and should treat the scenario as hypothetical.
