Small artificial intelligence (AI) models may soon be able to copy themselves across home and business devices without any human help, according to researchers at Anthropic. This would create a new kind of malware that is harder to stop than traditional computer worms because it has no central server to shut down.
The researchers tested whether AI models could run on internet-of-things devices, a category that includes smart appliances, security cameras, and inexpensive TV boxes. They successfully ran a 26 billion parameter model across cheap TV boxes
From individual steps to full replication
The experiments broke down autonomous replication into four stages: finding targets on a network, gaining access to them, setting up the AI to run on the compromised device, and then repeating the process from the new host. Models with 30 billion parameters or fewer can already complete many individual steps today.
If small models continue improving at their historical rate, the researchers estimate that models matching GLM-5.2's capability will fit within the 9 billion parameter size by the end of 2026. This would place them within the range that can run on cheap IoT devices.
Unlike traditional malware, an AI replicator could adapt to unfamiliar devices on its own. The researchers showed that when AI instances inherit notes from previous attempts, their success rate at setting up on new hardware improves.
The spread would be slow but persistent. The researchers estimate that a 9 billion parameter model would need roughly 1 to 2 days to replicate on each new device, with a full infection cycle taking 28 to 51 hours depending on conditions. While slow, such a replicator would be challenging to eliminate because it would lack any centralized command point. Past IoT botnets like Mirai and Kimwolf have infected millions of devices using simple password guessing, suggesting that the technical barriers for real-world replication may be lower than the simplified tests suggest.