AI news · Anthropic · security
Anthropic warns small models could soon spread onto gadgets
Anthropic researchers ran a mid-sized model across 15 cheap TV boxes and say self-copying AI on home devices may be months away.
Within months, small AI models could learn to install copies of themselves on cheap connected gadgets in your home, from smart ovens to no-brand Android TV boxes, Anthropic researchers wrote on Monday 12 October 2026. Think of a computer worm that brings its own brain along.
The team bought 15 off-brand TV boxes at about $55 each, linked them and ran Google’s Gemma4-26B-A4B model across all of them at 0.39 tokens (fragments of words) per second. Slow, but working. On 8 of 9 kinds of hardware, a 9-billion-parameter model (parameters measure a model’s size) managed to get a model running.
Models up to 30 billion parameters already handle many parts of the job, the post says. Putting all the parts together is where they still fail. On the current pace of progress, the team expects small models to finish the whole job by the end of 2026. With no single off switch, it would be hard to stop, Unite.AI notes. It would also spread slowly: the number of infected devices would take days to double.
Anthropic’s suggested defences sit with internet providers and device makers: spotting the traffic, quarantining infected networks and shipping safer default settings.
Sources
- Anthropic: AI models may be capable of self-replicating without access to a datacenter. Checked
- Unite.AI: Anthropic Red Team Says Small AI Models May Soon Replicate on IoT Devices. Checked
- Digit: Anthropic warns AI models may soon self-replicate without access to datacentres. Checked
Alexandre S.
Alexandre S. is a creative ad strategist and copywriter with ten years' experience in digital marketing across multiple channels.