In a published research post, the company said its experiments suggest that smaller AI models may be only months away from being capable enough to replicate across everyday devices, including smart ovens and inexpensive Android TV boxes.
The researchers warned that a self-replicating AI system could hijack poorly secured devices, use their computing power to run itself and continue spreading across networks. Unlike conventional malware hosted on a central server, such a system could have no single point that authorities or security teams could shut down.
Anthropic’s warning has drawn concern from some cybersecurity specialists, although others have questioned whether the company’s projections overstate the immediate danger.
Researchers warn of AI-powered botnets
Anthropic’s Frontier Red Team, which investigates potential security risks associated with AI, tested whether models could identify vulnerable devices, install themselves on them and continue spreading without human assistance.
The company found that models with 30 billion parameters or fewer could already perform many of the individual tasks required for replication, although they struggled to complete the entire process reliably. In simplified tests, more capable models successfully completed the sequence.
Read also : Bank of England warns AI debt could amplify financial risks
Anthropic also found that a model with nine billion parameters could run on eight of the nine hardware types tested.
The researchers said smaller models could potentially spread by compromising internet-connected devices, which often have weaker security protections than conventional computers. A model too large to run on a single device could also be divided across several machines, allowing them to work together.
Anthropic warned that an AI replicator could eventually seek access to more powerful computing resources, including data centres, or obtain computing capacity through stolen credentials and other means.
The company said that if smaller models continue improving at their historical rate, they could reach the capabilities required for autonomous replication by the end of 2026.
A familiar cyberthreat with new capabilities
Self-replicating malicious software is not new. Computer worms have been responsible for major cyber incidents, including the 2017 WannaCry ransomware attack, which disrupted systems around the world and locked users out of their files.
Anthropic’s concern is that AI could make such attacks more adaptable. Rather than relying entirely on instructions written in advance, an AI-powered worm could potentially identify unfamiliar devices and work out how to compromise them, allowing it to spread across a wider range of hardware.
The company called for stronger security measures to make connected devices harder to compromise and for preparations to contain self-replicating systems if they emerge.
