An AI program that copies itself from device to device has long been a science-fiction scenario. New research from Anthropic’s Frontier Red Team suggests it is quietly becoming an engineering question. In experiments published on 12 October, the team reports that small AI models are “months away” from being capable of self-replicating across ordinary consumer hardware — smart ovens, routers and cheap Android TV boxes included.
The full findings are set out in Anthropic’s research post, “AI models may be capable of self-replicating without access to a datacenter.”
Running on a pile of $55 TV boxes
To test whether a datacenter is really necessary, the researchers bought 15 off-brand Android TV boxes costing about $55 each and networked them together. Across that modest fleet they ran a 26-billion-parameter model, split into pieces so each box — with just 4 gigabytes of RAM — held a fraction of the whole. The improvised cluster worked, serving output at 0.39 tokens per second: painfully slow, but functional. The point was not speed. It was proof that capable AI inference no longer requires expensive, controllable infrastructure.
In separate tests, a 9-billion-parameter model succeeded in getting inference running on eight of the nine hardware types the team tried. Models of 30 billion parameters and under could already complete many of the individual steps of replication, though they still made mistakes when stringing those steps together end to end. In the team’s simplified evaluations, stronger models such as GLM-5.2 managed the complete sequence.
Why security teams should care
A self-replicating AI would be unusually hard to deal with. Unlike ordinary malware, it would have no central server to seize or shut down — Anthropic notes there would be “no centralized point to shut off.” And unlike a conventional computer worm with a fixed payload, an AI replicator could reason about unfamiliar devices it encounters and adapt on the spot.
There are real limits, at least for now. The researchers estimate any early replicator would spread slowly, with a doubling time measured in days rather than minutes, and today’s small models still fumble the full end-to-end process. But the trajectory is the concern: if small models keep improving at their historical rate, the team projects they will reach the required capability level by the end of 2026, with practical weight-based replicators possible soon after.
Part of a wider capability race
The study lands as AI capabilities — and worries about governing them — accelerate together, a tension we explored in our look at Gartner’s 2026 AI hype cycle. Anthropic’s red team is explicit that publishing the work is a defensive move: understanding how replication could happen is the first step to detecting and stopping it. For device makers, the message is blunter — the billions of cheap, poorly secured gadgets already plugged in around the world are, potentially, an AI habitat waiting to be occupied.