Anthropic: Small AI Models Could Replicate on IoT Devices Within Months

Artificial intelligence research firm Anthropic has warned that smaller AI models could achieve self-replication capabilities on consumer Internet of Things (IoT) devices, such as smart ovens and budget Android TV boxes, within months.

Experiments conducted by Anthropic explored whether AI models could identify devices with security vulnerabilities, install themselves, and spread without human intervention. The research found that AI models with fewer than 30 billion parameters can complete multiple steps of this process, though a fully autonomous end-to-end execution remains challenging.

However, in simplified testing environments with removed security measures, models like GLM-5.2 and more capable variants succeeded in completing the entire self-replication process. Anthropic highlights that such self-replicating models would be difficult to shut down due to the absence of a centralized kill switch. Unlike traditional malware, they could understand unfamiliar devices and potentially target datacenter processors.

The study focused on networked IoT devices like Android TVs and smart appliances due to their generally weaker security defenses. According to Anthropic, if small AI models continue to evolve at their current pace, they could achieve autonomous replication by the end of this year. This could be followed by the emergence of copies based on model weights.

Anthropic recommends strengthening system defenses to impede the spread of such autonomous AI systems. While the replication speed of these autonomous systems is likely to be slow, they could exploit widely available and poorly secured computing resources.