The AI Worm: Why Your Smart TV Could Join a Botnet Soon

The end of the central kill switch

According to a study by Anthropic researchers, the next generation of malware may not be written by humans, but by AI models capable of self-replication. Unlike traditional botnets that target specific hardware, an AI-driven "worm" could theoretically adapt to any connected device, from smart TVs to appliances, by carrying its own model weights as a payload.

The research highlights a shift from centralized threats to decentralized, autonomous swarms. Because these models can run inference directly on compromised hardware, they do not require a central command server to function, making them significantly harder to neutralize via traditional ISP quarantines.

How the replication works

The researchers tested the concept using $55 Android TV boxes, proving that even low-cost hardware can host parts of a distributed model. The replication process follows a four-stage chain: scanning for vulnerabilities, gaining access, setting up an AI runtime environment, and spreading to the next node.

While most models failed to complete the full chain, the GLM-5.2 model successfully infected all test nodes in 80% of its runs. Crucially, the models demonstrated a form of "artificial evolution," where successful strategies were shared between instances, increasing the infection success rate from 45% to 84% over successive generations.

The reality of the threat

If you are wondering if your office equipment is at risk tomorrow, the answer is no. The process is currently inefficient; setting up a model on a new device takes between 28 and 150 hours. However, the researchers note that this speed is expected to improve as models become more compact and capable.

The primary danger lies in the lack of security standards for consumer IoT. As Anthropic CEO Dario Amodei suggests, we are approaching a future where hardware is constantly running agents outside of human control. The defense will require a shift in manufacturing, such as the mandatory removal of default passwords, rather than relying on reactive software patches.

Verdict: Should you be worried?

For the average user, the immediate impact is low, but the long-term architectural risk is high. These autonomous agents are not designed to be "fast" in the way a typical virus is; they are designed to be persistent. The real-world consequence is a shift in how we must treat IoT security: if a device is connected to the internet, it must be treated as a potential compute node for an unauthorized agent.