Anthropic warns AI models could soon spread across smart devices on their own

Small AI models could soon spread without human intervention across internet-connected devices, from smart appliances to Android TVs, Anthropic has warned, raising fears that self-replicating systems could become increasingly difficult to contain.

Anthropic says new experiments suggest small AI models could be just months away from copying themselves onto everyday devices, including smart ovens and lesser-known Android TV boxes.

The company tested whether AI models could identify vulnerable devices, install copies of themselves and continue spreading without human involvement.

How close is AI to spreading independently?

The experiments found that smaller AI models, with fewer than 30 billion parameters, could already perform many of the steps required to replicate themselves, although they were not yet able to complete the entire process independently.

However, in simplified tests with some security protections removed, GLM-5.2 and more advanced models managed to complete the full replication process.

Anthropic warned that self-replicating AI systems could be especially difficult to contain, with no single switch capable of shutting them down.

Unlike traditional malware, such systems could adapt to unfamiliar devices without human guidance and potentially target processors in data centres.

Why everyday devices could be vulnerable

The research examined a range of internet-connected devices, including Android TVs, smart appliances and other gadgets that often have weaker security protections.

Anthropic warned that, if smaller AI models continue to improve at their current rate, they could become capable of spreading independently by the end of 2026, with the ability to copy their own model weights potentially emerging soon afterwards.

Anthropic calls for stronger security

Even if these capabilities emerge, Anthropic said the systems would probably spread relatively slowly, relying on widely available devices with weaker security.

The company urged stronger safeguards to prevent self-replicating AI from spreading across connected devices, warning of the growing security challenges posed by increasingly capable models.

Read more: 

• Anthropic researcher quits, warning AI 'could kill us all'
• Daily Brief: The latest in science and technology
• Anthropic CEO calls for slower pace of AI model development over fears of misuse

Tags

News Artificial Intelligence AI Anthropic AI security