Anthropic has issued a serious warning that small artificial intelligence models could soon become capable of self-replication and active propagation through internet-connected devices. While this scenario raises concerns among cybersecurity professionals, experts emphasize that creating a fully functional AI "worm" is still far off, and opinions regarding the urgency of this risk remain divided.
Exploring Distributed Capabilities
In October 2026, Anthropic's AI safety testing team published a study suggesting that self-replication capabilities could become a practical reality if smaller models continue to improve at their current pace. During the experiment, researchers deployed a model across 15 interconnected, low-cost Android TV receivers. The experiment demonstrated that computational resources from multiple common household devices can be successfully pooled to run AI, albeit at a low speed.
Potential Targets and Vulnerabilities
In the scenario modeled by researchers, potential targets could include home appliances and small business equipment, which typically feature computing power but weak system security. Such a distributed infrastructure would be difficult to shut down from a single management point. However, this remains a researched risk rather than a confirmed practice of a self-propagating AI network.Contradictory Data
The study has sparked mixed reactions and debate among experts. Cybersecurity outlet BleepingComputer emphasized that Anthropic's research team has not yet demonstrated a fully autonomous AI worm on a real Internet of Things device, noting that installation on weak hardware could take hours or days. Meanwhile, former UK National Cyber Security Centre head Ciaran Martin argued that the information was somewhat exaggerated, stating that running a model on cheap devices is insufficient to claim a security crisis. Conversely, Zhengshiong Li from the University of Colorado Denver expressed concern over AI's adaptability during device compromises.Analysis and Implications
These findings require a clear distinction between capabilities proven in a controlled test environment and the ability to self-replicate in real-world conditions. The expert community agrees that this should be treated as a scenario requiring further study rather than an immediate final forecast, highlighting the ongoing necessity to evaluate safety boundaries as autonomous technologies evolve.
