Anthropic has raised fresh concerns about self-replicating AI and connected devices. Its latest research suggests smaller models may soon copy themselves across everyday hardware.
The warning matters for telecoms and enterprise teams. Networks already carry traffic from millions of low-cost connected devices. If AI can spread through them, security teams may face a new containment challenge.
Anthropic said “Our experiments suggest that small AI models may be months away from being capable enough to self-replicate on commercial internet-of-things (IoT) devices, like smart ovens and off-brand Android TV boxes.”
The research focused on whether AI needs expensive data centres to operate. Researchers tested if models could run on cheaper consumer electronics instead. That shift could lower the barrier for experimentation. It could also widen the attack surface.
In one test, Anthropic ran an AI model across 15 Android TV boxes. Each device cost about $55. The setup ran slowly, but it still worked. That result showed basic consumer hardware can support distributed AI tasks.
However, speed was not the main concern. The real issue was autonomy. Researchers then checked whether AI could install and configure itself. They wanted to know if human guidance remained necessary.
One model configured itself on eight out of nine hardware types. Another copied itself across all eight test devices. It achieved this in four out of five attempts. Such results suggest rapid progress in model adaptability.
For security teams, that adaptability brings clear operational risks. A self-spreading system may not rely on one server. It could move across routers, media boxes, appliances, and other endpoints. That makes traditional shutdown methods less effective.
The company said: “A replicator like this would be challenging to shut down, since there’s no centralised point to shut off- though it may be slow (with a doubling time of days).”
At the same time, the findings do not mean a crisis has arrived. The tested systems were slow and limited. They also operated in controlled environments. Real-world networks add obstacles, including authentication, patching, monitoring, and device variation.
Still, telecom operators should watch this area closely. Many networks now support dense IoT deployments across homes, cities, and enterprises. Weak devices can become entry points into larger systems. Visibility and segmentation will become even more important.
The research also showed another notable trend. AI models improved when they received notes from earlier attempts. That means future systems could learn from failed installations. Over time, they may handle unfamiliar devices with greater accuracy.
This development could support useful applications in maintenance and automation. AI might configure edge devices or recover systems faster. Yet the same capability could help malicious tools spread.
Anthropic’s work adds urgency to AI security planning. Vendors, operators, and enterprises need stronger device controls. They also need clear monitoring for unusual model activity. The next wave of connected infrastructure must assume smarter threats.
