Anthropic's Red Team Built an AI Worm That Hid Across 15 TVs

By Elena Vasquez

Anthropic's offensive research team says it built a malicious AI program that spread autonomously between internet-connected televisions — and that real-world versions could work by the end of the year.

The finding comes from the company's red team, an internal group that stress-tests systems by attacking them. In a report published this week, the researchers described a self-replicating AI worm. It's a threat without human-like qualities or doomsday framing. It's a practical engineering problem, and that's exactly what makes it unsettling.

What did the researchers actually build?

The program targeted ordinary smart TVs. According to Anthropic, it moved through internet-connected devices on its own, with no human operator pushing it along at each step. That autonomous spread is what earns it the "worm" label, the same basic behavior that defined computer worms for decades.

The technical twist is how it handled power. AI systems usually need a fair amount of computing muscle to run. These TVs don't have it. So the researchers' program split parts of itself across 15 televisions, the report states, drawing a little processing power from each. The distributed program still functioned.

That detail matters more than it first appears. It means the hardware barrier that has quietly protected cheap, underpowered devices — the smart speaker in your kitchen, the TV in a hotel lobby — is not the wall security teams assumed it was.

Why does Anthropic think this arrives by year's end?

The company's assessment is blunt: successful AI worms need more capability on less powerful devices, and current trends make that feasible by the end of the year.

In other words, Anthropic isn't flagging a distant theoretical risk. It's pointing at a timeline measured in months. The same efficiency gains that let small models run on phones and laptops also lower the bar for malicious code that must squeeze into weak hardware.

Why this cut through the noise?

Security experts are drowning in AI warnings. As the original reporting noted, alerts about what AI might do in the future — alongside alarming accounts of what it has already done — now arrive so fast that even professionals struggle to triage them.

This report stood out for the opposite reason. No sentience claims. No apocalyptic speculation. Just a red team demonstrating that autonomous, self-distributing AI code can survive on hardware nobody thought could host it.

The stakes sit squarely in the consumer device market. Smart TVs, and by extension other ordinary connected devices, ship in the hundreds of millions and rarely receive serious security attention. Anthropic's researchers said their program could do the same inside those other devices too.

The so-what is simple and close to home. If the capability curve holds, defenders won't be chasing malware that tricks users — they'll be chasing code that quietly parcellizes itself across the cheapest screens in the house.

Anthropic's warning reframes the smart-device security problem at exactly the moment efficiency gains make it real.