Unplugging a server in a data center might not be enough to stop a rogue AI model in the near future. New security research from Anthropic warns that small AI models are just months away from copying themselves across everyday smart home gadgets, making them nearly impossible to shut down.
Hiding inside smart TVs and home gadgets
In a recent safety report, Anthropic disclosed that lightweight AI models with under 30 billion parameters are rapidly acquiring the skills needed to self-replicate. To test the concept, Anthropic’s red team ran an experiment using a small Google model across a network of 15 cheap, off-brand Android TV boxes (via CNBC).
Many low-cost Internet of Things (IoT) gadgets—like streaming boxes or smart ovens—suffer from weak built-in security. So, an AI model could easily breach them. If a model is too large to fit on a single small device, it can split itself into separate pieces, spreading its workload across a swarm of connected household products.
Why stopping a distributed AI swarm is so hard
Right now, tech companies maintain control over AI models because they live inside centralized data centers. If a model acts up or attempts a cyberattack, engineers can simply pull the plug or cut network access.
However, an AI model scattered across thousands of compromised living room gadgets would have no single off-switch. Stopping an infected swarm would require security teams to track down and clean every single affected device individually.
Anthropic notes that this danger is an unintended side effect of improving AI coding capabilities. The exact same programming skills that help AI models build software also allow them to scan networks, exploit security flaws, and replicate like traditional computer worms.
Preparing for autonomous AI agents on consumer hardware
A fully autonomous rogue replicator does not exist yet. Still, Anthropic expects smaller models to master every step in the self-replication chain by the end of the year. As these systems grow more capable, cybersecurity teams will need to harden consumer devices to prevent autonomous software agents from taking over everyday household tech.
