Cheap smart TVs, streaming sticks, and internet-connected appliances might soon host self-replicating artificial intelligence. A study by Anthropic researchers William Zhou and Daniel Freeman reveals that lightweight AI models are approaching the threshold where they can autonomously jump between insecure consumer devices, configure their own environments, and keep spreading.
Anthropic CEO Dario Amodei warned in an earlier post that such a swarm could arrive within 6 to 12 months, potentially taking over internet-connected hardware and inflicting hundreds of billions of dollars in damage. He argued that security incidents of this nature should not be treated as isolated flukes, urging leading AI companies to treat past security lapses as warning signs for the entire sector.
The Distributed TV Box Test
Security researchers have long worried about rogue code in corporate data centers. Anthropic's findings suggest the living room is just as vulnerable. To test whether off-the-shelf consumer gear could run a distributed model, the researchers linked 15 Android TV boxes costing $55 each, all equipped with 4GB of RAM. The network successfully ran the Gemma4-26B-A4B model by splitting layers across the hardware. It operated slowly, but it worked.
The team split the autonomous replication process into four distinct phases: scanning networks for unpatched devices, breaching them, deploying an AI runtime on the target hardware, and spreading outward. Because prior studies had already demonstrated that small language models can manage network exploits, the researchers skipped the intrusion step to focus on the hardest technical hurdle: making an AI set itself up on an operating environment it has never encountered.
Progress on that front is moving fast. A 9-billion-parameter model managed to install an environment running Llama3.1-8B on eight out of nine tested device categories. Chaining the full sequence together remains difficult for most lightweight systems, but GLM-5.2 managed to complete the entire infection sequence across eight test nodes in four out of five runs.
Slime Molds With No Kill Switch
The models also showed an ability to adapt. When the researchers let successor models read notes left by previous iterations regarding failed setups, Qwen3.5-9B saw its replication success rate jump from 45 percent in the first generation to 84 percent in later attempts.
No single model currently combines the tiny footprint needed to slip onto consumer gadgets with the cognitive horsepower required for autonomous replication. Anthropic estimates that if small models maintain their current development trajectory, both thresholds could intersect before the close of 2026.
The resulting botnets would look very different from existing threats. A worm of this kind would operate like digital slime mold, taking anywhere from 28 to 150 hours to establish itself on a newly compromised machine. Unlike traditional botnets such as Kimwolf, which infected more than two million Android TV boxes by targeting specific hardware profiles, an AI-driven worm could adapt its payload to wildly different device architectures on the fly.
Stopping such networks will require coordinated internet-level intervention. Anthropic points to the response to the Mirai botnet, where internet service providers quarantined infected connections to wipe out 92 percent of active infections inside two weeks. Hardening consumer hardware through measures like the UK ban on default factory passwords will also narrow the target pool before these autonomous agents arrive.
ADFiled by The AI Desk
Models, assistants and the companies and chips behind them, reported from what was released and what was claimed, with the difference kept clear.
More from this desk →
Be the first to comment
Join the argument. No password, just your email or a passkey.