Anthropic Ran an AI Across 15 Cheap TV Boxes. The Next Step Is the Scary Part.
Fifteen Android TV boxes. Roughly $55 each. One AI model spread across them.
It sounds like the beginning of a science-fiction horror story. But the experiment is real, and Anthropic's Frontier Red Team says the implications deserve attention. In research published October 12, the company showed that an AI model could run across a collection of inexpensive connected devices, while separately testing whether other models could install inference software and replicate through simplified networks.
There is one crucial distinction: Anthropic did not release a self-replicating AI worm into the wild, and the 15 TV boxes did not independently infect other TVs. The concern is what happens if future small models become good enough to join those steps together.
The most surprising experiment involved hardware from the bargain bin
The researchers connected 15 inexpensive Android TV boxes, each with limited memory, and split a model called Gemma4-26B-A4B between them. Each box held part of the network. Together, they produced responses at about 0.39 tokens per second.
That is painfully slow by chatbot standards. It also demonstrates something important: a useful AI model no longer necessarily needs one powerful GPU or a centralized data center to run. With distributed inference, many weak devices can collectively host it.
Connected TVs, streaming boxes, cameras and other consumer hardware are attractive targets because they can be numerous, inexpensive, poorly updated and always switched on. This is a question about the security of the internet of things as much as the future of AI.
Did the AI copy itself from TV to TV?
No. The widely shared TV-box demonstration established distributed inference, not a self-propagating infection.
Anthropic also designed separate, controlled evaluations of a replication chain: identifying available systems, obtaining execution, installing the necessary software and initiating another copy of the process. Some more powerful models completed simplified end-to-end replication tasks, while smaller models succeeded at individual pieces but often failed when required to combine them reliably.
In another evaluation, a 9-billion-parameter model got inference running on eight of nine tested hardware types. That is meaningful evidence of increasingly flexible software setup, but it is not proof that the model can autonomously invade unknown home networks.
Anyone calling this an actual household AI outbreak is skipping the evidence.
Why would an AI worm be different from ordinary malware?
Traditional computer worms can already spread automatically. They do not need AI. The difference researchers are worried about is adaptability.
Instead of relying on a fixed sequence of instructions, a future agent might inspect an unfamiliar device, work around a broken dependency, adjust its installation approach and choose a different route after an error. That makes defensive assumptions harder: the device may be weak, but the instructions attacking it need not be fixed.
There is still an enormous gap between a controlled model-installation benchmark and a resilient real-world campaign. Running inference is slow on weak hardware. Transferring model weights requires bandwidth and storage. Limited context and reliability compound errors. Some traditional malware remains far cheaper and more effective.
Anthropic's warning is about a trajectory, not a demonstrated apocalypse.
The detail that should worry device makers
The most actionable conclusion is not to fear your television becoming conscious. It is that weak, forgotten hardware creates places where unwanted software can survive.
A box that rarely gets security updates, exposes remote-management services or reuses default credentials can already become part of a botnet. Adding better autonomous software could eventually make such networks more adaptable.
Organizations should start with mundane defenses: inventory connected devices, disable unnecessary remote access, update firmware, isolate untrusted IoT hardware from sensitive networks and investigate unusual traffic. None of these protections requires believing that an AI superworm already exists.
What happens next?
Anthropic argues that continued capability gains could make smaller models better at end-to-end replication within months. That is a forecast conditional on future progress. Other security experts question both the timeline and how much practical advantage AI would bring over existing malware.
The right question for the next experiments is narrow: Can one small model, on unfamiliar physical devices, complete the entire process reliably without researchers simplifying the task? Until that happens, the most dramatic headlines remain ahead of the evidence.
The Zerionia view
The genuinely strange part is not that Anthropic built a virus. It did not demonstrate one in the wild.
It is that a pile of cheap consumer boxes can host a distributed AI model at all, while agent software is becoming increasingly capable of configuring unfamiliar systems. That combination gives cybersecurity a new threat model to prepare for. It also gives the internet another reason to stop treating bargain-bin connected hardware as harmless.
