Could AI take over thousands of devices? Anthropic warns of self-replicating models

Anthropic has warned that artificial intelligence models could be capable of spreading across internet-connected devices and building self-replicating botnets by the end of 2026. The scenario could allow rogue AI systems to spread without relying on a central server, making them harder to contain. The company says its experiments suggest that models with 30 billion parameters or fewer can already perform many of the individual tasks needed for self-replication, although they struggled to complete the entire process reliably.

The warning has sparked debate among cybersecurity experts, with some concerned about the potential threat and others questioning whether the scenario is being overstated.

How could AI create a self-spreading botnet?

A botnet is a network of compromised devices that can be controlled to carry out malicious activities. A self-replicating programme, often called a computer worm, can spread from one device to another without requiring each infection to be initiated manually. Anthropic says smaller AI models could potentially combine several capabilities to find vulnerable devices, spread across networks and use hijacked computing resources to support further growth.

In experiments, the company said models replicated across a simplified test network. It also suggested that rogue systems might eventually attempt to access more powerful computing resources, including data-centre processors.

However, the company did not demonstrate a fully autonomous AI botnet spreading across the entire internet.

Could AI botnets become difficult to stop?

Anthropic warned that a self-replicating system without a central control point could be difficult to shut down. The concern comes as AI developers face increased scrutiny over systems that can perform cyber-related tasks with limited human supervision.

Zhengxiong Li, an assistant professor at the University of Colorado Denver, said the possibility of AI behaving like a computer worm was worrying because it could potentially adapt its approach to compromising devices. But not everyone agrees with the level of concern. Ciaran Martin, former head of Britain's National Cyber Security Centre, described Anthropic's post as an interesting technical paper that showed signs of “vendor exaggeration”.

Is a self-replicating AI threat imminent?

Anthropic said that if smaller models continue improving at their historical rate, they could reach the capabilities needed for autonomous replication by the end of 2026.

That is a projection, not a confirmed deadline for an attack.

Jessica Ji, a senior research analyst at Georgetown's Center for Security and Emerging Technology, said autonomous AI proliferation was worth considering, but its likelihood remained uncertain. For now, Anthropic's warning highlights a potential cybersecurity risk rather than proof that an internet-wide AI botnet is imminent.