Anthropic has issued a warning that small artificial intelligence (AI) models could develop the ability to easily replicate themselves via internet-connected devices within months, raising concerns regarding cybersecurity and the possible risks of rapidly autonomous technology.
As per the Claude AI manufacturer, such models could possibly infiltrate devices such as smart ovens and Android TV boxes, and use their computing power to operate and spread to other connected systems without depending on expensive data centres.
Anthropic cautioned that shutting down a self-replicating AI-powered system could become quite difficult as it may spread across several devices without a major point of control.
The company addressed a fully rogue, self-replicating AI-centric system that does not yet exist.
However, it stated AI agents are rapidly capable to complete individual steps involved in the replication process and passing adaptations to future copies.
As part of its research, Anthropic tested a small Google AI model, Gemma4-26B-A4B, across a network of 15 inexpensive, off-brand Android TV boxes. The company noted that devices with weak security could provide potential entry points for malicious AI systems.
Moreover, Anthropic warned that models that are too large to run on a single device could divide into smaller components, known as “shards,” and run across multiple devices.
As per the company, a worst-case scenario may involve an entire network of several AI models that work on different hardware and using multiple methods to replicate themselves.
Potential threats could include cyberattacks, unauthorised access to computing resources and the theft of application programming interface (API) keys.
Anthropic said future AI replication could take several forms, ranging from slow, distributed systems operating on insecure devices to more powerful models capable of breaching protected computer systems.
As per Anthropic, future AI replication could take multiple forms, ranging from slow, distributed systems operating on weak devices to high-end models that are able to breach protected computer systems.
The warning comes amid increasing industry concerns regarding AI safety as models become highly capable. However, Anthropic stressed that the accurate form and timeline of autonomous AI replication remain uncertain.
