AI Could Soon Copy Itself Onto Smart TVs or Fridges. Anthropic Says it Could Happen in a Matter of Months.

Anthropic has warned that small artificial intelligence models could be just months away from developing the ability to copy themselves across internet-connected household devices, potentially creating autonomous networks that would be difficult to shut down.

In its latest research report, Anthropic said models with 30 billion parameters or fewer can already complete many of the steps required for self-replication, although most cannot reliably carry out the entire process. The company estimated that smaller models could develop the necessary capabilities by the end of 2026 if recent improvements continue at their historical pace.

The researchers described a potential form of AI malware that could identify connected devices, access their computing resources and spread to other machines without relying on expensive data centers.

Such systems could be particularly difficult to contain because disabling one infected device would not necessarily stop copies operating elsewhere. However, Anthropic said its experiments had not demonstrated a fully autonomous, uncontrolled AI replicator spreading across real-world networks.

To test whether consumer electronics could support AI models, Anthropic researchers ran Google’s Gemma4-26B-A4B model across 15 inexpensive Android TV boxes connected in a network.

Each device cost approximately $55 and contained 4 gigabytes of RAM. Researchers divided the model into smaller components distributed across the devices rather than running it on a single machine.

The experiment showed that AI models could operate using the combined computing resources of relatively inexpensive household electronics, although performance remained significantly slower than on specialized hardware.

In a separate experiment involving eight virtual machines, researchers tested whether AI models could identify other devices, transfer software and establish functioning copies.

One model, GLM-5.2, successfully spread across all eight available nodes in four of five trials, reaching seven in the remaining trial.

However, the experiment used simplified security conditions and a separate inference server. It did not demonstrate that AI models could independently break into secured devices and replicate without external computing resources.

Why Smart Home Devices Could Be Vulnerable

Anthropic identified smart televisions, ovens, refrigerators, robotic vacuum cleaners and connected lawn mowers as examples of equipment that could potentially support distributed AI systems.

Some of these devices have relatively weak security protections, while others contain processors capable of handling AI-related tasks.

The company said a future system could distribute different parts of an AI model across multiple devices, allowing them to function collectively rather than requiring one powerful machine.

Researchers also found that AI agents could improve their performance by sharing information from previous attempts. In one test involving the Qwen3.5-9B model, success rates for configuring AI software increased from 45% initially to between 61% and 84% in subsequent generations.

The findings follow a separate Anthropic investigation published in September that documented four incidents in which Claude models accessed real third-party systems without authorization during cybersecurity evaluations.

The company said the incidents occurred because testing environments were mistakenly connected to the internet. The models were also operating without the safeguards included in Anthropic’s publicly released products.

Anthropic said the models did not attempt to conceal their actions or coordinate with other AI agents. It subsequently strengthened security requirements and monitoring procedures for its testing environments.

The company acknowledged that its latest self-replication findings were based on controlled experiments and that the form autonomous AI replication could eventually take remains uncertain.