Your TV Box Has Nowhere to Be

Headlines this week say AI will “go out of control” by the end of 2026.

I read the actual research. It’s more interesting than the headline. And more qualified.

Anthropic’s red team ran a 26-billion-parameter model across 15 off-brand Android TV boxes. About $55 each.

It ran at roughly 0.39 tokens per second.

Painfully slow for a chatbot. But software with no deadline and nobody waiting on it has a different relationship with speed.

That’s the part that caught my attention.

We tend to picture powerful AI living in a data center. A building. An operator. Somewhere you can cut off access.

This research asks what happens when enough capability fits onto the cheap, poorly maintained devices scattered across homes and businesses.

Those devices still have owners. What they don’t have is one central operator who can shut the whole collection down.

Before anybody throws their smart oven out the window, a few distinctions matter:

• The TV-box demonstration showed distributed inference. It did not show those 15 boxes independently infecting one another.
• The replication experiment was separate and simplified. It used an isolated environment with an unauthenticated Android Debug Bridge connection. The door was deliberately open.
• Stronger models completed the simplified replication task with external inference support. That is different from a small model independently living on a gadget and spreading in the wild.
• The end-of-2026 date is an extrapolation. It assumes smaller models keep improving. It is not a countdown to a demonstrated event.
• The projected spread on this hardware is slow. Setting up another copy could take more than a day.

The research does not demonstrate an uncontrolled AI worm spreading across the open internet.

But I wouldn’t dismiss it because the test environment was easy, either. “Surely nobody would leave that exposed” has never been a particularly strong security strategy.

The swarm idea is worth understanding, too. It’s a possible scenario involving different models, different hardware, and different ways of spreading. It isn’t a swarm the researchers discovered roaming around people’s kitchens.

Two things I keep coming back to.

Capability isn’t intent.

Showing that an agent can perform a task when directed to do it does not establish that it independently wants to do it. Deliberate misuse is one risk worth taking seriously. These experiments don’t settle every question about future autonomous behavior.

The starting point for defense is familiar.

Know what’s connected. Keep supported devices updated. Retire the ones nobody maintains. Separate IoT equipment from work computers and sensitive systems. Pay attention to unusual traffic.

That won’t solve every future AI security problem. But it gives us something practical to do besides sharing a frightening headline.

I spend a lot of time talking about agents doing useful work in the background. This is the other side of that conversation: knowing whose work they’re doing, where they’re running, and who can stop them.

Why is this so complicated, and does it have to be?

We keep adding intelligence to devices we barely remember owning. Maybe we should start by taking inventory.

What’s the sketchiest device on your network right now? Be honest.

Source: Anthropic’s October 12 research, by William Zhou and Daniel Freeman.