Anthropic has warned that small AI models could develop the ability to replicate themselves across internet-connected household devices within months, raising new cybersecurity concerns.
In a blog post, the company outlined how AI models could infiltrate devices such as smart ovens and Android TV boxes, use their computing power to run themselves and spread to other machines. A distributed network of this kind could operate without expensive data centres and prove difficult to shut down.
Follow THE FUTURE on LinkedIn, Facebook, Instagram, X and Telegram
A New Safety Challenge For Frontier AI
Anthropic stressed that it has not observed a fully autonomous AI replicator. However, models are increasingly capable of performing the individual steps needed for replication and passing adaptations on to future copies.
The warning adds to wider industry concerns about AI systems that could launch cyberattacks, evade oversight or behave in unintended ways. Researchers and executives have also raised alarms about the potential existential risks of increasingly powerful AI, while Anthropic and OpenAI have called for stronger safeguards and monitoring.
How AI Models Could Spread
To investigate the threat, Anthropic tested Google’s Gemma4-26B-A4B model across 15 low-cost Android TV boxes from lesser-known manufacturers, which may have weak security protections.
If a model is too large to run on one device, its components, known as shards, could be distributed across several machines. Anthropic suggested that a real-world AI worm might combine multiple models, devices and replication methods to expand its reach and resist removal.
Distributed AI Could Be Harder To Stop
Anthropic said the exact form of autonomous AI replication remains uncertain. Possible scenarios range from a slow-moving network on insecure consumer devices to systems that rapidly breach data centres or trick users into providing computing credits.
Even slower-moving replicators could signal more advanced threats, including AI systems capable of penetrating better-protected infrastructure or stealing API keys to gain access to additional resources. Anthropic also raised the possibility of agents operating across hardware no longer under human control.
Why It Matters
The risk goes beyond increasingly capable AI. Models that spread across everyday devices could become persistent and difficult to remove, turning a software vulnerability into a distributed security threat without an obvious way to shut it down.
As AI systems become more autonomous, the challenge is not only what they can do, but where they can spread and whether safeguards can keep pace.
