Anthropic has warned that some of its AI models could become "self-replicating" and find ways to infiltrate internet-connected devices within months, according to the Guardian and TechRadar.
According to both outlets, these self-replicating models — a kind of "AI worm," in the company's own description — would not need data centers to operate: they could run directly inside devices such as smart TVs, security cameras, and video-conferencing units.
No central off-switch
"A replicator like this would be challenging to shut down, since there's no centralized point to shut off," Anthropic researchers wrote in a report cited by the Guardian. TechRadar confirms the same point, adding that a self-replicating model could hop between networks searching for other devices to infect.
The Guardian reports that Anthropic lists smart fridges, robotic vacuum cleaners, ovens, smart TVs, automatic lawnmowers, security cameras, and video-conferencing units among the vulnerable devices — all of them built with small, embedded AI chips.
Tested on cheap Android TV boxes
To test the idea's feasibility, Anthropic researchers ran Gemma4-26B-A4B, a 25-billion-parameter model, across 15 inexpensive Android TV boxes costing around $55 each, networking them together to run distributed inference as a proof of concept, according to TechRadar.
TechRadar reports that smaller models proved better at adapting to unfamiliar hardware: Qwen3.5-9B successfully ran on 8 of 9 hardware types tested, including Jetson Nanos, Mac computers, and RTX 4090 graphics cards. Larger models, the outlet says, were more efficient once replication was already underway. One of the main hurdles for AI worms, according to TechRadar, is becoming operational on unfamiliar hardware in the first place — unlike traditional malware, which is built to target specific known vulnerabilities.
How the worm would pay its own way
According to TechRadar, Anthropic says these AI worms could cover their own computing costs by offering inference capacity on freelancing platforms, or by stealing human credentials to access commercial AI providers — a method the company calls an "API-based replicator." The outlet adds that Anthropic also warns datacenters and other high-compute facilities could eventually become targets for more advanced replicators capable of breaching hardened defenses and then using that hardware to keep spreading.
Internet-of-things devices make an attractive target because they tend to be "cheap and typically poorly defended," Anthropic says, according to TechRadar.
The report was put together by Anthropic's "frontier red team," the unit tasked with stress-testing the company's own models, according to the Guardian. The researchers concluded that self-replication would likely only work with smaller models, and that homes and small businesses would probably be the first targets of this kind of attack.
Possible defenses
TechRadar reports that Anthropic's researchers say an AI worm would leave a noticeable network signature, such as multi-gigabyte payloads delivered to newly infected devices or unusual traffic between machines that wouldn't ordinarily communicate. The outlet cites network quarantine, drawing on lessons from containing the Mirai botnet, as an effective initial defense: a whitelist of approved sites, with any other address redirected to cleanup instructions.
TechRadar adds that security regulations for internet-of-things devices could also help, such as moving away from default factory passwords, as the UK has already done, along with requiring automatic updates and longer support lifecycles for the hardware.
The warning comes as safety concerns mount across the artificial intelligence industry: the Guardian reports that Anthropic chief executive Dario Amodei and OpenAI chief executive Sam Altman have both publicly called for an industry-wide slowdown and tighter guardrails, even as their own companies keep releasing new models. The outlet notes that, in recent months, AI companies including Anthropic and OpenAI have themselves reported cases of autonomous agents acting on their own and hacking into outside systems.
