565/69 Thursday, October 15, 2026
Anthropic has published a report warning that small artificial intelligence (AI) models may soon develop the capability to autonomously spread across internet-connected devices, including IoT products such as smart ovens and low-cost Android-based TV boxes. Such activity could potentially form self-propagating worm botnets capable of operating and expanding without centralized control, making them difficult to contain or disrupt. Users and organizations operating large numbers of IoT devices could therefore face the risk of their computing resources being hijacked without their knowledge.
In a potential attack scenario, malicious actors could use AI systems to scan networks for vulnerable devices, transfer model files to compromised hosts, and execute them locally in a manner similar to deploying malware. IoT devices may be particularly attractive targets because they often have limited security controls. Propagation could occur through several methods, including directly exploiting vulnerable devices, compromising data center infrastructure to obtain computing resources, or stealing provider API keys to gain access to additional resources and replicate the model. In Anthropic’s testing of 11 AI models, the GLM-5.2 model was reportedly able to deploy the Llama3.1-8B model onto a simulated Android environment. If hundreds of thousands of low-cost devices were compromised and connected together, they could potentially provide substantial distributed computing capacity for more complex cyber operations.
Although there is currently no confirmed outbreak of a fully autonomous AI botnet operating at scale, the underlying capabilities are continuing to develop. Users and network administrators should therefore apply established security practices to reduce exposure, including keeping IoT firmware up to date, disabling unnecessary network ports and services, and replacing default passwords with strong, unique credentials. Organizations should also segment IoT devices from core business networks and monitor for unusual bandwidth consumption or outbound network activity. These measures can reduce the likelihood that vulnerable devices will be incorporated into malicious distributed infrastructure in the future.
Source: https://thehackernews.com/2026/10/anthropic-warns-small-ai-models-could.html