Anthropic warns AI worms could spread across smart devices as cybersecurity risks grow

Anthropic has warned that advances in artificial intelligence could make self-replicating malware more practical, potentially allowing malicious AI systems to spread across internet-connected devices such as smart TVs, cameras and household appliances.

The warning follows experiments by the company’s Frontier Red Team, which investigates how advanced AI systems could be misused. In a report published on 12 October, researchers examined whether AI models could help create a new kind of computer worm that moves between devices, adapts to different hardware and runs without relying on a central data centre.

The research does not show that a small AI model can already carry out this entire process reliably in the real world. Instead, it identifies technical capabilities that could bring such a threat closer as models improve.

What makes an AI worm different?

Traditional computer worms spread by exploiting security weaknesses, often copying malicious code from one vulnerable machine to another. An AI-powered worm could add another layer: using a model to reason through unfamiliar systems, adapt its approach, and help configure itself for different hardware.

Anthropic calls one potential design a weight-based replicator. Rather than relying solely on conventional malware code, it could carry or retrieve AI model weights and establish the software needed to run the model on a compromised device.

That flexibility could make attacks more adaptable. A system encountering an unfamiliar device might attempt to work out how to run compatible software instead of depending entirely on a fixed set of instructions written for a specific platform.

The concern is particularly relevant to internet-connected devices, many of which receive security updates less frequently than computers and smartphones. Devices left with known vulnerabilities could provide opportunities for malware to spread.

What Anthropic's experiments actually demonstrated

Anthropic tested whether AI models could configure inference environments across different hardware platforms and coordinate replication across a network.

In one experiment, researchers used a large model to orchestrate a simplified replication process across eight cloud-hosted Android emulators. The test relied on an exposed Android Debug Bridge interface, rather than demonstrating a successful attack against a representative range of secured consumer devices.

Separate tests explored whether smaller models could establish inference environments on different hardware. The researchers found that models could complete individual steps, but smaller models still struggled to perform the entire replication process reliably from beginning to end.

That limitation matters. Establishing an AI model on a device, finding a vulnerable target, gaining access, configuring the software, and successfully spreading to another machine are separate technical challenges. Demonstrating some of these steps does not mean an autonomous, self-propagating AI worm is ready for widespread deployment.

Why the threat could become harder to contain

Anthropic argues that improvements in smaller AI models could reduce the gap between today's experimental capabilities and more practical attacks. Its researchers suggested that models with around 30 billion parameters or fewer could reach the necessary level of capability by the end of 2026 if historical progress continues. This is a projection, not a confirmed deadline.

A successful worm operating across many independent devices could also be harder to disable than malware that depends on one central server. Security teams might need to identify and clean up infections across different networks and hardware platforms rather than shutting down a single control point.

However, practical obstacles remain substantial. Running AI models on inexpensive devices can be slow, transferring model weights consumes bandwidth, and the process may generate activity that security monitoring systems can detect. Researchers and outside experts also disagree about how quickly these scenarios could become realistic.

What users and manufacturers should do

The immediate response does not require treating every smart appliance as an imminent AI threat. The established security measures remain important: install firmware updates, replace devices that no longer receive security patches, change default passwords, and disable remote-access features that are not needed.

Manufacturers also have a role in providing longer security-support periods, reducing unnecessary network exposure and making it easier to update devices. For businesses, network segmentation and monitoring can help limit the damage if one connected device is compromised.

The bottom line

Anthropic's research highlights a credible area of emerging cybersecurity risk: AI could make malware more adaptable and help automate tasks needed for self-replication. But the experiments have not established that small models can already spread rapidly and independently across ordinary, properly secured consumer devices. The priority is to strengthen device security now while continuing to test these capabilities and distinguish demonstrated risks from worst-case projections.