Anthropic’s Frontier Red Team has warned that small AI models could soon copy themselves across consumer electronics. This could create decentralised networks of computing power that need no dedicated data centres.
The team published its report on October 12. Its experiments suggest that small AI models could gain the ability to replicate on their own on commercial Internet of Things (IoT) devices within the next few months. Such a system could exploit poorly secured devices and install the software needed to run an AI model. It could then spread to other machines without constant human help.
Running a Large Model on Cheap Hardware
The researchers ran a 26-billion-parameter model, Gemma4-26B-A4B, across 15 inexpensive, off-brand Android TV boxes. Each box had 4GB of RAM. The model ran at about 0.39 tokens per second.
The test shows that work usually done on specialised servers can be split across cheap consumer devices. The devices pass data to one another to complete the job together.
Spreading Is the Harder Part
Running a model across devices is only part of the task. An autonomous AI worm would also need to find targets, gain access and transfer its payload. It would then have to set up inference on the new hardware and repeat the cycle.
In a separate controlled test, the team used Google Cloud Platform nodes running Android 15 emulators. The setup simulated a replication environment. It checked whether a model could hop between machines through unauthenticated debugging ports and start inference on a target.
Anthropic reported that GLM-5.2 and stronger models could complete the full replication process in this simplified test. The experiment does not show that an uncontrolled AI worm has spread across consumer devices in the real world.
Adapting to Unfamiliar Hardware
The team also tested whether models can set up inference on unfamiliar hardware. Qwen3.5-9B succeeded on eight of the nine hardware types tested. This suggests AI systems could adjust their software to different environments, instead of following a fixed installation routine.
That could make future AI worms more adaptable than traditional malware, which relies on set exploits and instructions. An AI-driven system could troubleshoot failed installs, change its approach and keep trying across different devices.
Why Containment Could Be Hard
A self-replicating AI system would not need data centre GPUs to survive, Anthropic warned. Smart TVs, ovens, robot vacuums and other connected appliances could supply the computing power to run parts of a model.
The report describes a threat model in which the model carries its own weights as part of the malware payload. Weights are the numerical parameters needed for inference. If a model is too large for one device, its weights could be split across several machines.
This design could make containment harder. The network would have no central server that defenders could shut down. Anthropic expects such replication to be relatively slow. The network could take days to double in size, instead of spreading at once.
A Timeline and a Call to Defenders
Anthropic found that models with 30 billion parameters or fewer can already complete many of the individual steps needed for autonomous replication. They still make mistakes when they join the steps into a full process.
The researchers looked at how fast smaller open-weight models have improved. They projected that models able to complete the whole process could appear by the end of 2026.
Anthropic said defenders will need to strengthen security across consumer and enterprise devices. They should also prepare for systems that may be hard to remove once they spread.
