Anthropic tested a lightweight AI model across 15 smart TV boxes, each with 4 gigabytes of RAM. The test points to a possible foothold for AI worms in ordinary home devices, but it doesn’t show an autonomous worm spreading through real-world smart homes.
In a report published Monday, Anthropic warned that AI worms might soon break into WiFi-enabled TVs, cameras, and other devices, then use them to replicate and spread elsewhere. The company said the threat could be mere months away if AI models continue improving at their current rate.
A worm is malware that copies itself and spreads by scanning a network for other devices. Anthropic researchers are examining whether rogue AI agents may soon gain that ability, particularly on everyday hardware connected to home networks. Examples include smart TVs, fridges, lights, robovacuums, security cameras, and toilets.
Many households could be vulnerable, the report says, because devices may come from dubious manufacturers and users may not have properly secured their networks. Consumer devices also increasingly contain more powerful chips, which could make it feasible for a small AI model to infiltrate and run on them.
Datacenter GPUs would be an obvious high-value target for malware, Anthropic says. But the report also notes, “IoT devices – which are cheap and typically poorly defended – could also provide enough compute to run a capable model.”
The report cites previous research suggesting that a potential AI worm could carry its model weights as it replicates, then run inference on hardware it takes over. On lighter-weight home devices, such a worm would likely spread slowly, more like a “digital slime mold.” Anthropic’s test ran a lightweight model across 15 smart TV boxes with 4 gigabytes of RAM each.
Researchers also found that AI worms could figure out unfamiliar devices without help, unlike non-AI malware. That ability could help an agent work through the varied hardware it encounters, though it’s only one part of the process required for self-replication.
Anthropic’s report distinguishes those demonstrated steps from a complete threat: “We don’t yet have a full rogue replicator, but agents are becoming increasingly capable of completing the individual steps in the replication chain.” For now, the practical concern is that poorly secured IoT devices could offer a foothold, not that AI worms have been shown infecting smart homes in the wild.
