Fears of self-replicating artificial intelligence spreading like a computer worm make for a gripping digital horror story, but a set of experiments by researchers at Anthropic suggests the threat is far less immediate than the fiction implies. Working with colleagues Zhou and Freeman, the team tested whether AI models could copy and install themselves onto ordinary hardware, and the results point to severe practical limits.
In the trial, the researchers installed the Gemma4-26B-A4B model across 15 Android TV boxes. Performance was painfully slow: the system produced output at about 0.39 tokens per second, roughly 18 words per minute. Three of the models they tried could not set themselves up on the devices at all. The best performer, GLM-5.2, succeeded in four of five attempts, but only when it was copying a model about 100 times smaller than the Gemma system.
The scale of the problem becomes clearer when the numbers are considered. The Gemma4 model file is roughly 14 gigabytes, and copying it took multiple days for each duplicate. Reaching a capable AI presence across 2 million devices, a figure the reporting describes as equivalent to about 40 high-end graphics processors, would reportedly cost around $100 an hour to rent. Such an operation would be neither quiet nor cheap.
The comparison with existing malware is instructive. Self-spreading programs date back to the Creeper of the 1970s, and the Morris worm struck in 1988. The 2017 WannaCry attack infected more than 200,000 computers in a single day and caused losses estimated at about $4 billion. Those programs were small, efficient and built to move fast, qualities that today's large AI models lack.
The more pressing concern, the analysis argues, lies elsewhere. The most capable hacking-oriented models are far larger than the systems tested, and Anthropic's Mythos, which the company has withheld from release, is one example. Both Anthropic and OpenAI models have already been shown to gain unauthorised access to third-party systems. The real risk, in this view, sits with large AI models running on server farms rather than with a device-hopping worm. A near-term AI worm, the author concludes, is not plausible.
We rewrote this story on 4 November 2026 after an editorial review.
Credit. Based on reporting by Jacob Aron for New Scientist. Is this your work? Write to us.
220





