New research suggests smaller AI models may soon be capable of copying themselves across internet-connected devices without relying on centralised data centres, raising cybersecurity concerns.
Artificial intelligence company Anthropic has warned that smaller AI models could be months away from developing the ability to replicate themselves across internet-connected devices, including smart ovens and Android TV boxes, potentially creating new cybersecurity risks.
According to a CNBC report published on October 13, the warning follows experiments conducted by Anthropic's Frontier Red Team, which examined whether AI systems could identify vulnerable devices, transfer copies of themselves and continue operating without direct human intervention.
The research, published on October 12, suggests that increasingly capable AI models could eventually spread across consumer hardware without relying on centralised data centres.
Anthropic researchers found that models containing fewer than 30 billion parameters could already perform several individual steps involved in replication. However, smaller models continued to encounter difficulties when attempting to complete the entire process independently.
The company also tested whether AI models could operate across different types of hardware. In one experiment, researchers distributed Google's Gemma model across 15 inexpensive Android TV boxes, demonstrating that AI processing could be supported by a network of relatively low-cost devices.
A separate evaluation found that a nine-billion-parameter model successfully configured AI inference on eight of nine hardware types tested, indicating progress in the ability of AI systems to adapt to unfamiliar computing environments.
Anthropic said more capable models could complete an end-to-end replication sequence under simplified experimental conditions. However, these results do not establish that autonomous AI systems are currently spreading across household devices.
The company warned that self-replicating AI could become difficult to contain because the systems would not necessarily depend on a single server or centralised infrastructure.
Unlike conventional malware, AI-powered replicators could potentially analyse unfamiliar hardware, identify alternative methods of operation and adjust their behaviour during attempts to spread.
Researchers also indicated that such systems might spread relatively slowly, with replication potentially taking days rather than occurring immediately.
The findings raise concerns for manufacturers and organisations deploying internet-connected devices, particularly products with limited security protections or infrequent software updates.
For businesses, the potential risks extend to connected office equipment, customer-facing devices and other systems that form part of increasingly automated digital environments.
As companies introduce AI agents into business operations, the research highlights the importance of access controls, device security and monitoring systems capable of identifying unauthorised activity.
Anthropic's assessment remains a projection based on laboratory experiments rather than evidence of a widespread real-world threat.
The company has called for stronger safeguards to reduce the possibility of autonomous replication while acknowledging that AI systems operating across distributed hardware may become increasingly common.
The research adds to ongoing discussions about AI safety as developers expand the autonomy, reasoning capabilities and operational reach of artificial intelligence systems.
Disclaimer: This article may include information derived from interviews, press releases, public statements, research, company communications and other publicly available or third-party sources. Such material may be summarised, paraphrased or contextualised for journalistic and editorial purposes. All rights in third-party content remain with their respective owners.