Anthropic study warns that small AI models could soon replicate across everyday devices
decoder ·
A new study by Anthropic researchers warns that small AI models may soon be capable of autonomously replicating across poorly secured consumer devices like smart TVs and IoT appliances. In tests using inexpensive Android TV boxes, the GLM-5.2 model successfully completed a full replication chain across all eight test nodes, demonstrating the technical feasibility of such an 'AI worm.' While no current model is both small enough and capable enough to execute this attack in the wild, researchers predict this gap will close by the end of 2026. The proposed malware would carry its own weights, allowing it to configure itself on unfamiliar hardware without human intervention, making it decentralized and difficult to shut down. Anthropic CEO Dario Amodei has warned that such a swarm could potentially take over the internet within six to twelve months, causing hundreds of billions of dollars in damage. Although defense strategies like ISP quarantine and banning default passwords exist, the AI worms are expected to adapt to these measures autonomously. The researchers compare the slow but persistent nature of these replicators to a digital slime mold, highlighting the challenge of finding a central kill switch.
decoder