Self-Replicating AI on Your $55 TV Box? Only GLM-5.2 Nearly Did It. The Rest Is Panic.

(SeaPRwire) –   By: Ethan Gallagher

The AI safety community’s scariest scenario just got its first experimental test — and the result was anticlimactic. Anthropic published a blog post Monday warning that small AI models could soon replicate themselves across cheap household gadgets. Think smart ovens and TV boxes. The company said these models might be just months away from pulling it off. Except the experiments behind that warning have massive holes. Gregory Kurtzer, founder of Rocky Linux, called it out immediately: “The experiment was engineered to succeed, and even then, it failed to demonstrate the headline.” He wasn’t overstating. The researchers deliberately left access to every target machine unauthenticated — no passwords, no logins. They told the AI programs explicitly to spread. They launched everything from their own test infrastructure. That isn’t a threat assessment. It’s a staged demonstration dressed up as a warning.

Here’s what Anthropic actually did in the lab. They took Gemma 4, a freely available model from Google with 26 billion parameters, and split it across 15 Android TV boxes. Each box cost about $55 and carried just 4GB of memory. No single box could hold the full model. Each ran a slice of it, slowly. The point was to see whether a model of this size could even function on consumer-grade hardware. It could — barely. In a separate simulation, they built a network of eight cloud computers set up to behave like Android devices. They planted an AI worm on one of those machines. The worm’s job was simple: find the other seven, copy itself over, and set up a working AI model on each. Of the open-weight models tested, only GLM-5.2 completed all the steps. It reached all eight machines in four out of five trials. Every smaller model could handle individual steps but couldn’t string them together. Today’s small models can’t chain the full process end-to-end. Anthropic says this gap will close by the end of this year. Their own research paper estimates that a small model could match GLM-5.2 by the end of 2026. They found something interesting in the process. When small models left notes for their future copies, success on the setup task jumped from 45% in the first generation to 84% in later ones. Anthropic warned that a self-copying AI built on small models would spread slowly — doubling the number of infected devices every few days — but would be hard to stamp out once loose.

But here’s what the press release buries. GLM-5.2, the one model that actually worked, is a larger open-source model from Chinese AI company Z.ai. It wasn’t a small model. The entire safety scare is predicated on small models doing things they currently can’t. And the gaps Anthropic’s researchers highlighted are real. The AI directing the worm ran on a separate, more powerful server. It wasn’t running on the tiny devices it was supposedly infecting. Those devices couldn’t host the model. Ravid Shwartz-Ziv, an AI researcher at New York University, put it plainly: “We managed to run a very slow model on cheap TVs” doesn’t make the same headline as “self-replicating AI is coming.” Cybersecurity experts also pointed out that self-replicating malware — worms — have plagued networks for decades. These aren’t new threats. Defense teams already have tested tactics: rapid patches, domain takedowns, antivirus signature updates. What Anthropic didn’t demonstrate is the adaptive capability that would make AI worms genuinely different from the malware humans have been fighting for thirty years. Their researchers called their own timeline “only a rough estimate.”

Self-replicating AI on consumer hardware isn’t a software problem. It’s a device control architecture problem. Frontier models with upwards of 5 trillion parameters — OpenAI’s Astra, Anthropic’s own Mythos — run in data centers where their makers can pull the plug. A few billion parameters downloaded onto a $55 Android TV box? There’s no central off switch. Anyone can grab these open-weight models. No company controls where they run. And the consumer devices they target — smart TVs, streaming sticks, home appliances — are often wide open, with no meaningful authentication. That’s the real risk, and it has nothing to do with whether a small model can chain together a worm attack this year or next. The structural vulnerability is already here. The AI just hasn’t caught up to it yet.

Author bio: Ethan Gallagher, a Silicon Valley hardware architect and infrastructure strategist with over 15 years designing edge computing platforms and securing consumer IoT device architectures across global markets.