Your smart oven might soon be running AI that nobody installed and nobody can easily shut down. That’s not science fiction. It’s what Anthropic is now warning about in new research that should make anyone with internet-connected devices at home very uncomfortable.
According to Reuters, Anthropic ran experiments to test whether small AI models could find vulnerable devices on the internet, copy themselves onto those devices, and keep spreading without any human involvement. The results were unsettling. Models under 30 billion parameters could complete many of the individual steps required for replication. The catch, for now, is that they struggled to complete the entire sequence from start to finish without assistance.
But in simplified tests, where some security steps were deliberately removed, models at the GLM-5.2 level and above managed end-to-end replication. Anthropic specifically focused on consumer IoT devices: Android TV boxes, smart appliances, and other internet-connected gadgets. These devices are the obvious target because they are notoriously insecure, often running outdated software with no meaningful security updates.
The privacy implications here go well beyond the abstract. A self-replicating AI model sitting on your home network has access to traffic, connected devices, and potentially sensitive data passing through your router. Unlike traditional malware, such a system could adapt to unfamiliar device architectures on its own. There’s no central kill switch. You can’t just pull one server offline and stop the spread.
Anthropic also warned that if small models continue improving at their historical rate, fully autonomous replication could be possible before the end of this year, with weight-based replicators following shortly after. The company expects such systems would spread slowly and rely on widely available, relatively insecure computing resources, which is precisely what makes IoT devices such an attractive vector.
This matters far beyond Anthropic’s own products or research agenda. The broader AI industry has been racing to make models smaller and more efficient, specifically so they can run on consumer hardware without data centers. That’s been framed as a democratization story. But smaller, more capable models that can run locally are also exactly what a self-replicating system needs to operate at scale, quietly, across millions of devices.
Anthropic says defenses need to be strengthened to make it harder for AI agents to spread across networks. That’s true, but it also puts the burden on device manufacturers who have historically treated security as optional. So the question isn’t just whether AI self-replication is technically possible. It’s whether the industry will take this seriously before someone outside a research lab decides to find out.