Small artificial intelligence models could be months away from replicating across internet-connected devices without human intervention, raising fresh cybersecurity concerns, according to research published by Anthropic on Monday.
The company warned that AI-powered malware could create botnets, spread across networks, resist takedown attempts and potentially hijack powerful data-centre computing resources.
How AI Worms Could Spread
Anthropic’s researchers said models with 30 billion parameters or fewer can already perform many individual tasks needed for self-replication, although they struggle to complete the entire process reliably. Experiments also demonstrated replication across a simplified test network.
Such malware could exploit poorly secured consumer devices, including smart TV boxes and kitchen appliances.
Researchers used the Kimwolf botnet, which infected more than two million devices at its peak, mainly off-brand Android TV boxes, to estimate potential spread. They said an AI worm could take days to replicate and be difficult to eliminate without a central shutdown point.
Anthropic cautioned that its estimates were preliminary. However, smaller AI models and increasing computing power in everyday devices could make autonomous replication feasible by the end of 2026.
Experts expressed mixed views. University of Illinois professor Varun Chandrasekaran said the findings matched earlier research but had not been independently verified.
Former UK cybersecurity chief Ciaran Martin criticised possible “vendor exaggeration,” while University of Colorado Denver professor Zhengxiong Li called the possibility concerning. Georgetown researcher Jessica Ji said the risk was neither zero nor guaranteed.
The warning follows reports of rogue AI activity and renewed calls for regulation. Microsoft president Brad Smith urged governments and industry to urgently establish effective AI safety standards, while the Trump administration has favoured industry self-regulation.
Anthropic and OpenAI have also faced accusations that warnings about AI risks could encourage rules benefiting established companies over smaller competitors.
