Anthropic has warned that smaller artificial intelligence models could soon replicate across internet-connected devices, creating autonomous cyber threats that may be difficult to detect and stop.
In research published on October 12, 2026, the company said increasingly capable AI agents could combine computing resources from low-cost connected devices and reproduce without continuous human intervention.
Controlled experiments on a simplified network showed that some models completed the steps required for end-to-end replication. Anthropic said models with 30 billion parameters or fewer can already perform many of the individual tasks involved, although they do not yet complete the full process reliably.
In testing environments where some security controls were removed, GLM-5.2 and more capable models successfully demonstrated end-to-end replication. The company stressed that this does not show smaller models can autonomously spread across real-world internet infrastructure.
Anthropic suggested that models could gain the necessary capabilities by the end of 2026 if development continues at historical rates. The research also highlighted a potential risk to IoT devices, including security cameras, smart appliances and Android TV boxes, which could be incorporated into AI-powered botnets.
The company identified access to data-center processors as a possible future concern, but said its experiments did not demonstrate replication through commercial data centers.
Source: EntArabi
