A TV that starts thinking for itself without being asked sounds like the setup for a bad sci-fi movie. That's roughly what Anthropic tried to recreate in a lab last week, and the outcome was more unsettling than even the researchers expected.
The company ran an AI model across 15 cheap Android TV boxes — the kind of generic streaming dongles sold for $20 on marketplaces. The model didn't just run on them: it fully replicated itself, completing the entire chain from copying its own weights to launching on a new device and adapting to its quirks. Separately, the team showed that larger models are already capable of "end-to-end" replication across a controlled network.
Among the open-weight models tested, only GLM-5.2 — a large model from China's Z.ai — completed the full cycle. Smaller models could handle individual steps but couldn't chain them together. That's why Anthropic isn't describing a live threat but a horizon: by its estimate, a model small enough to run on consumer hardware and smart enough to fully replicate could emerge before the end of 2026.
The difference from a conventional computer worm is fundamental. Classic malware answers to a command server — take that down, and the infection stops spreading. A system built on autonomous AI agents, researchers wrote, has "no centralized point to shut off." Each copy figures out how to behave on a new device on its own, tries different approaches, and passes successful adaptations to the next copy. The pace isn't alarming by itself — doubling the number of infected devices would take days, not seconds — but there's no kill switch beyond patching the security holes in the devices themselves.
And those holes are plentiful. Smart ovens, robot vacuums and security cameras have long drawn criticism for weak default passwords and rare firmware updates — exactly why Anthropic flags them as the likeliest entry point. No "wild" replicator has been found so far; this is a lab experiment, not an incident. But publishing this kind of worst-case red-teaming of its own technology fits a pattern for Anthropic, which has previously urged the industry to slow the race to build ever-more-capable frontier models.
