Technology

Could AI Soon Copy Itself Onto Your TV?

Martin HollowayPublished 4w ago3 min readBased on 9 sources
Reading level
Could AI Soon Copy Itself Onto Your TV?
source:anthropic.com

Anthropic warned that small AI models could be months away from gaining the ability to copy themselves across cheap devices. The Verge

The warning moves attention away from the data center. The largest models, possibly with trillions of settings that control how they behave, need huge computers that keep them in data centers. Smaller models can run on normal home hardware and are often free to download.

To test that case, Anthropic researchers split Google's Gemma 4, a model with 26 billion settings, across 15 Android smart TV boxes. No box could run it alone. Like neighbors each holding a few pages of a long manual, the boxes worked together to run the full model.

A second test looked at spread. Anthropic built a simulated network of computers set up to act like Android devices. It tested an AI-powered worm told to find new machines, copy the AI model onto them, and start it on each one.

Only one model tested, GLM-5.2 from China's Z.ai, spread across the whole simulated network. Smaller models finished single steps of self-copying even when they did not spread across the whole network. Anthropic researchers predict smaller models could match GLM-5.2 by the end of the year based on how fast they are improving.

The test split the attacker from the package. The model that was copied in the worm test was Meta's Llama 3.1-8B, with 8 billion settings. It was different from and much smaller than the model running the attack. In short, a stronger agent handled finding, breaking in, and installing, while a small, easy-to-move model was the thing being copied.

Anthropic published its research on Oct. 12, 2026. Anthropic The company stated that small AI models may be able to self-replicate across cheap connected devices like Android TV boxes with no datacenter.

Coverage on Oct. 13, 2026 described the same concern in practical terms. Anthropic described AI models potentially getting into connected devices such as smart ovens and Android TV boxes. CNBC Other reports described models that could clone themselves and infect connected devices such as smart TVs and appliances, CBS wreak havoc on the web, Reuters and become challenging to shut down once spread across smart home devices. Yahoo Anthropic also flagged a danger of botnets, or networks of hacked devices, automating cyberattacks in connection with self-replicating AI.

Separately, Anthropic announced an expanded Cyber Verification Program on Oct. 6, 2026. The program gives advanced cyber tools and looser safety blocks to approved security professionals.

The broader context here is containment. Defenders have assumed that cutting off central computers limits misuse. Running across many cheap devices breaks that assumption. Persistence does not need stealing a large model or holding a data center. It needs enough hacked low-end devices to host a useful small model and enough skill to keep finding the next device.

For practitioners, the split that matters most is between the coordinator and the package. An 8-billion-size package is already portable. The limit is whether the agent can handle different devices, passwords, setup differences, and recovery from partial failure. Anthropic's result suggests that limit is shrinking.

In my view, the defense problem looks less like model control and more like basic smart-device hygiene at scale. Slow patching, default passwords, movement between home devices, and poor visibility into strange network traffic are familiar issues. What changes is automation of the whole loop, from finding to installing to running, without human pacing.

One point worth flagging for security teams is the testing approach Anthropic points to. Giving checked researchers stronger cyber tools mirrors how powerful security tools have long been shared in controlled ways. Defenders need real tools to build detection. Wider access raises misuse risk. Anthropic has chosen a gated program for that access.

The optimistic read is still available. The same traits that make small models easy to spread make defensive models easy to spread. Local checking, on-device warning signs, and fast triage do not need data-center scale if a small model can run across pooled home hardware. The limit will be installing and maintaining it, not raw ability.