AI models may soon self-replicate without access to data centers, Anthropic warns
- Published On Oct 13, 2026 at 01:02 PM IST
Anthropic, in its latest blog post on Monday, warned that small AI models could become capable of replicating themselves across internet-connected devices, such as smart ovens and Android TV boxes, within the next few months.
The company cautioned that such AI systems could be difficult to shut down because they would operate across multiple devices without a central point of control.
Although their spread could be slow, taking days to double, they could independently adapt to unfamiliar devices, unlike conventional malware.
Advt
In its tests, Anthropic found that a 9-billion-parameter AI model successfully ran inference on eight of the nine types of hardware evaluated.
The replication and adaptation could occur through AI-powered worms, which can spread across devices without human intervention.
Anthropic explained that a worm is a computer program that copies itself from one device to another. Previous research has shown that AI agents can behave similarly by scanning networks, exploiting security vulnerabilities and transferring AI model weights to other machines within a GPU cluster.
The AI platform outlined three potential ways AI worms could spread: by running smaller models across connected smart devices, using larger models to target data centres and steal computing resources, or accessing AI services through stolen API credentials.
It warned that a more serious threat could emerge from multiple AI models working together across different devices and using various methods to replicate, making them harder to detect and shut down.
- Published On Oct 13, 2026 at 01:02 PM IST