Anthropic says small AI models may be approaching a capability long associated with science fiction and malware research: copying themselves across low-cost, internet-connected devices without relying on a traditional data center.
In research published October 12, the company’s Frontier Red Team tested whether relatively small models could set up inference on unfamiliar hardware and chain together the steps needed for autonomous replication. Anthropic’s conclusion was cautious but notable: the individual pieces are increasingly possible, even though the full system remains unreliable.
One experiment distributed a 26-billion-parameter Gemma model across 15 inexpensive Android TV boxes that cost about $55 each. The cluster ran at roughly 0.39 tokens per second. That is slow, but the point was not performance. It was to test whether useful model inference could exist outside conventional accelerator hardware.
Anthropic also reported that a 9-billion-parameter model successfully set up inference on eight of nine hardware types it tested. Models with 30 billion parameters or fewer can already complete many individual replication tasks, the researchers said, while stronger models such as GLM-5.2 completed an end-to-end replication sequence in a simplified, firewalled evaluation.
The research does not show an AI worm currently spreading through homes, smart TVs or appliances. Anthropic’s end-to-end tests used controlled environments, and the company explicitly says today’s smaller models still make mistakes when the separate replication steps are combined.
The concern is what happens if model capability keeps improving while compact open-weight systems continue getting easier to run on edge hardware. Anthropic estimated that a small model could reach the capability level used in its simplified replication test by the end of 2026 if recent capability trends continue. The company described that as a rough projection rather than a firm forecast.
A self-replicating model would resemble a computer worm, but with one important difference: the AI could potentially adapt to unfamiliar devices instead of relying entirely on fixed exploit logic. Anthropic says that could make decentralized infections harder to remove because there may be no single command server to shut down.
That distinction matters because conventional botnets already exploit weakly secured IoT hardware. BitcoinVersus.Tech recently covered PoeLLM turning AI servers into a cryptomining botnet, while earlier coverage looked at AI-assisted cyberattacks and defenses and the broader layers of cybersecurity.
Reuters reported on October 13 that some cybersecurity researchers consider the scenario plausible but not inevitable. The biggest open questions remain the same ones Anthropic’s own tests expose: how quickly a model could spread, whether it could obtain reliable execution on real devices, and whether the compute available on compromised hardware would be useful enough to justify the attack.
For now, the headline is not that an autonomous AI worm has escaped onto the internet. It is that researchers can increasingly test pieces of that threat model with real hardware, small models and ordinary networked devices. That makes IoT security, patching, default passwords and device isolation more relevant as AI capability moves closer to the edge.
BitcoinVersus.Tech Editor’s Note:
We volunteer daily to ensure the credibility of the information on this platform is Verifiably True. If you would like to support to help further secure the integrity of our research initiatives, please donate using the QR code provided
For best QR-code scanning results, please scan this donation code from a desktop or larger screen. Mobile screens may prevent camera recognition.
BitcoinVersus.tech is not a financial advisor. This media platform reports on financial subjects purely for informational purposes.
Leave a Reply