Anthropic Warns of Chilling AI Risk That Could Emerge Soon

Amazon AMZN backed Anthropic is warning that smaller artificial-intelligence models could be only months away from gaining the ability to replicate themselves across commercial internet-connected devices, raising a new class of cybersecurity and AI-safety concerns.

The company said models may soon be capable enough to spread across devices such as smart ovens and inexpensive Android TV boxes without relying on a central server.

That distributed structure could make an AI worm significantly harder to shut down.

Anthropic estimates that replication would likely take days rather than seconds or minutes, but the risk comes from persistence: once distributed across many devices, disabling one machine would not necessarily eliminate the system.

The company tested whether AI models could identify devices on a network, copy themselves onto additional machines, provision the resources needed to run and continue spreading without human intervention.

In one experiment, Anthropic ran an AI model across 15 Android TV boxes costing roughly $55 each, showing that a network of relatively low-powered devices could collectively support a larger model.

In another simulated-network test, one model successfully replicated across eight devices in most trials.

There is an important limitation.

Anthropic simplified the process of gaining access to those devices, meaning the tests do not prove that current AI systems could spread as easily across real-world networks.

Still, the company found that models with 30 billion parameters or fewer could already complete many of the individual steps required for autonomous replication, even though they struggled to combine those steps reliably.

If smaller models continue improving at their historical pace, Anthropic believes they could reach the necessary capability by the end of 2026.

Investor Takeaway

I think the bigger story is that AI safety risk may increasingly move from giant frontier models into smaller models cheap enough to run almost anywhere.

That changes the economics of the threat.

A system no longer necessarily needs an enormous cloud data center if smaller models can coordinate computation across thousands of inexpensive connected devices.

For Anthropic and the broader AI industry, this increases pressure for stronger safeguards around device security, model access, autonomous actions and containment.

It could also create opportunities for cybersecurity companies if AI-driven attacks become more autonomous and distributed.

The key development to watch is whether small models can move from completing individual replication steps to reliably chaining the entire process together.

Anthropic's research suggests that gap still exists. But its warning is that the gap may not remain open for long.