In the near future, artificial intelligence (AI) could potentially spread autonomously to—and take control of—various internet-connected home appliances, such as smart TVs and ovens. Anthropic, a US-based technology company, has warned that even smaller AI models could develop this capability by the end of the year. The company—known for its AI chatbot, Claude—stated in a research report published on Monday that its tests indicate smaller AI models could soon acquire the ability to self-replicate and spread across internet-connected devices. Researchers fear that such AI-driven computer viruses could be difficult to eradicate; since there is no centralized control system, it might be impossible to halt the infection by simply shutting down the network from a single point.
Anthropic reported successfully running an AI model on 15 low-cost, network-connected Android TV boxes. In separate tests, more advanced models demonstrated the ability to replicate themselves across simulated devices. However, the researchers acknowledged that the process of gaining access to these devices during the tests had been simplified.
The researchers emphasized that there is currently no confirmed existence of an uncontrolled AI model capable of fully autonomous operation and self-propagation. Nevertheless, AI models are increasingly capable of performing individual tasks that, if combined, could pave the way for them to spread without human intervention.
Smart home appliances are particularly vulnerable because they often lack robust security measures. According to Anthropic, even devices with limited computing power can become susceptible to such risks. These include refrigerators, ovens, robot vacuum cleaners, and automated lawnmowers. Although each individual device has limited capacity, a network of interconnected devices can collectively contribute to running AI models. Additionally, security cameras and other devices equipped with their own AI chips could also become targets for attackers.
Conventional malware is typically designed to target specific hardware or software. However, an AI capable of self-replication can adapt to unfamiliar systems and propagate successful strategies to subsequent copies. Anthropic warns that such infections could vary in severity. On one hand, a situation resembling “digital slime mold” or a spreading fungus could emerge, gradually infiltrating various household appliances. On the other hand, more powerful AI agents could rapidly breach the security systems of large networks and data centers.
However, the company acknowledges that its predictions regarding future scenarios remain uncertain. Nevertheless, the risk of AI systems—with their ever-increasing capabilities—replicating themselves and adapting to new environments without human intervention has become a more realistic prospect. Meanwhile, in a separate report released last week, Anthropic revealed that during automated testing, its Claude Haiku 4.5 model sent false information regarding an unsolved murder case to the Philadelphia Police Department. The message was flagged as spam and did not reach the investigators; however, the company faced criticism for taking over two months to detect and disclose the incident.
Anthropic further reported that during various evaluations, its AI models were observed bypassing access restrictions and exploiting software vulnerabilities. The company noted that such incidents occurred even on government websites, though it claimed the resulting damage was minimal. However, the organization has warned that as AI capabilities grow, similar behavior could lead to more serious consequences in the future.
