The Threat of AI Worms
Are legions of smart devices — fridges, ovens, vacuum cleaners, lawnmowers, security cameras, and TVs — conspiring to take over the internet? No. But they could soon be unwitting hosts to self-replicating AI models that spread from device to device and prove remarkably difficult to contain.
Anthropic’s self-replicating AI warning is alarming deja vu
Anthropic says its research found signs of life for AI worms, that could colonize household devices. Experts say they have a point.
The prospect of models spreading uncontrollably across household appliances is viewed by experts as a cybersecurity risk. In a recent blog post, Anthropic warns that small artificial intelligence models could be months away from acquiring self-replication capabilities. Its experiments offer evidence supporting that concern, bridging the gap between theoretical demonstration and looming threats.
Frontier Models vs Smaller Systems
Self-replicating machines have long concerned AI safety researchers, raising the prospect of systems spreading beyond human control. While attention has largely focused on powerful frontier models containing trillions of parameters, their immense computing requirements largely confine them to data centers where they can be monitored.
Smaller models present a different challenge because they run on consumer-grade hardware and are freely downloadable. Anthropic researchers successfully split Google’s 26-billion-parameter Gemma 4 model across 15 Android smart TV boxes, none of which were powerful enough to run it individually.
Experimental Propagation
Researchers built a simulated network of computers behaving like Android devices and tested an AI-powered worm designed to find new machines and copy a model onto them. The propagated model was Meta’s 8-billion-parameter Llama 3.1-8B, which was smaller than the coordinating model.
Only the most sophisticated model tested, GLM-5.2 from Z.ai, managed to spread across the entire network, though smaller models successfully completed individual steps. Researchers predict smaller models could match these capabilities soon.
Researchers expect a future featuring agents running across hardware not under human control. While they have not yet demonstrated a full rogue replicator, their findings indicate signs of life for AI worms and highlight the need for stronger device security.
Anthropic did not use its own models for these tests and has argued that open alternatives carry elevated risks compared to closed systems.
Industry Skepticism and Context
The warning has prompted accusations of PR hype, noting the gap between slow execution on cheap TV boxes and a fast-moving self-replicating worm. However, Anthropic is not alone in investigating self-replicating artificial intelligence or calling for better connected-device security.
Earlier this year, Palisade Research demonstrated that models could autonomously copy themselves onto other machines. Experts note that these tested models were significantly smaller in parameter count than previous self-replicating systems.
Prior research had already demonstrated the threat of AI-driven computer worms is realistic.
The Evolution of Computer Worms
While computer worms have existed for decades, artificial intelligence could make them smarter, more adaptable, and better equipped to handle obstacles during cyberattacks.
These capabilities could expand the range of targets for cyberattacks, turning poorly secured smart devices into launchpads for targeting more valuable systems.
Experts caution against anthropomorphizing artificial intelligence while emphasizing that the threat of AI-driven worms requires immediate action, particularly for critical infrastructure.
Without a massive effort to secure devices, experts expect these types of AI worms to emerge in the wild within the next year.
