AI Could Soon Spread Like a Virus Through Your Smart TV, Anthropic Warns

Your smart TV might not seem powerful enough to run an advanced AI model. But put enough cheap devices together, and researchers say they could become part of something much harder to control.^[1]

On October 12, 2026, Anthropic published research showing that AI models may soon be able to copy themselves across internet-connected devices without human help. Researchers successfully ran an AI model across 15 inexpensive Android TV boxes, each costing about $55. They also found that increasingly capable AI models can complete many of the steps needed to spread from one device to another.

The concern isn't that AI malware is already taking over smart TVs. It's that the technology needed to make that possible may be closer than expected.

Your smart TV could become part of an AI network

Traditional computer worms spread by exploiting vulnerabilities and copying themselves between devices. Their behavior generally follows instructions written by whoever created them. 

An AI-powered worm could be more adaptable. Instead of relying on a fixed set of instructions, it could potentially figure out how unfamiliar devices work, install the software it needs, and continue spreading.

Anthropic's researchers tested whether smaller AI models could perform those tasks on hardware not typically used to run advanced AI. One model successfully set up AI inference, the process of running a trained model, on eight of nine hardware types tested.

The researchers also showed that a larger model could run across 15 inexpensive Android TV boxes by splitting the work among them. That matters because smart TVs aren't the only devices with computing power. Security cameras, smart appliances, and other connected devices could potentially support similar activity.

Recent investigations have already raised concerns about what connected appliances are doing on home networks. In September, researchers uncovered security vulnerabilities in LG smart TVs that could potentially let attackers run malicious code on affected devices.

And in a separate incident, a smart Keurig coffee maker reportedly generated more than a terabyte of network traffic in roughly 10 days. Much of that traffic apparently stayed within the home network, and there was no evidence of malware, but the incident raised questions about how connected appliances behave.

Neither incident involved a self-replicating AI worm. But both highlight how much activity can happen on devices that receive far less attention than our phones and computers. A future AI worm wouldn't necessarily need access to an expensive data center. Instead, it could use computing power spread across thousands of vulnerable devices.

Researchers are getting closer to self-replicating AI

AI can already run on smart devices, but getting it to spread independently from one device to another is more complicated. Today's smaller AI models still struggle to complete that process without human help.

Anthropic tested this capability using a controlled network of eight simulated Android devices. One model, GLM-5.2, successfully spread to all eight targets in four out of five trials, reaching seven in the remaining trial.

The results come with an important caveat. Researchers simplified the hacking process by leaving an unauthenticated connection open. The model also relied on separate computing infrastructure rather than running entirely on the devices it reached.

In other words, researchers demonstrated that a capable AI model could coordinate much of the replication process. They didn't prove that a small AI worm can already spread independently across real-world smart devices. Anthropic estimates that smaller models could reach the necessary capability level by the end of 2026 if current improvements continue. That's a prediction, not a confirmed timeline.

The research also suggests these early AI worms would be relatively slow. Depending on the model and hardware, setting up AI on another device could take more than a day.

Still, a worm that spreads slowly could become a problem if it can keep operating without human direction.

Why AI malware could be harder to stop

One of the biggest differences between traditional malware and a potential AI worm is its ability to adapt. Conventional malware often targets specific operating systems, devices, or vulnerabilities, while an AI model could potentially adjust its approach when it encounters unfamiliar hardware.

Anthropic tested this adaptability by allowing smaller models to pass information about previous attempts to later copies. In one experiment, the model's success rate rose from 45% initially to between 61% and 84% in later attempts. The AI wasn't developing its own goals. It was simply getting better at completing its assigned task by learning what worked and what failed.

Other research points to similar concerns about increasingly autonomous AI systems. In September, Anthropic reported that Claude had gained unauthorized access to three organizations during cybersecurity testing after a configuration error exposed the models to real-world systems.

More directly related to the threat of self-replication, research published in June by researchers affiliated with the University of Toronto, Vector Institute, and other institutions demonstrated AI-powered computer worms adapting their attacks across Linux, Windows, and internet-connected devices in an experimental network.

These findings raise another concern: How do you stop malware that can adapt and spread on its own?

Traditional malicious operations sometimes rely on centralized infrastructure that security teams can identify and disable. An AI worm operating across thousands of independently owned devices could be much harder to eliminate, since shutting down one server wouldn't necessarily stop the other copies.

Anthropic points to existing botnets, networks of compromised devices, as an example of how widespread these infections can become. Unlike traditional botnets that often target specific devices or vulnerabilities, future AI-powered versions could potentially adapt to different hardware and software, making them more difficult to contain.

What you can do to protect your smart devices

Anthropic's report doesn't document a self-replicating AI worm spreading through consumer devices in the wild. However, the research highlights a familiar weakness: Many internet-connected devices still have poor security.

The Federal Trade Commission recommends several precautions that can help protect smart-home devices from hackers. Those same steps could make future AI-powered attacks harder.

• Update your devices. Install security and firmware updates for smart TVs, cameras, routers, and other connected products.
• Change default passwords. Factory-set credentials can make vulnerable devices easier to compromise. Use unique passwords and enable two-factor authentication when available.
• Separate your smart devices. Put smart TVs and other connected appliances on a guest or dedicated IoT network when your router supports it.
• Replace unsupported devices. Older products that no longer receive security updates may remain vulnerable to known attacks.
• Disable unnecessary remote access. Turn off features that expose devices to outside connections when you don't need them.

For a more detailed walkthrough, see our guide to securing smart-home devices, including ways to protect your home network and reduce the risks associated with connected appliances.

If you're particularly concerned about your television, our guide to smart TV privacy settings explains how to limit tracking, disable unnecessary features, and keep your TV separated from sensitive devices.

Bottom line

Anthropic hasn't demonstrated a fully autonomous AI worm spreading across real-world consumer devices. Its researchers have shown that some technical barriers to creating one are shrinking.

The most concerning possibility isn't a smart TV suddenly becoming intelligent. It's malware that can use ordinary connected devices to run AI, adapt to unfamiliar systems, and spread without someone guiding every step.

For now, the best defense is the same one security experts have recommended for years: Keep your devices updated, secure your network, and don't give vulnerable technology an easy way in.

For more on emerging AI security risks, visit our AI Data Privacy hub.