The world is buzzing with smart devices—gadgets that can connect to the internet and transmit data between one another. There are smartphones, smart speakers, smart TVs, smart refrigerators, smart rings, smart watches, smart vibrators (not joking), smart thermostats, smart ovens, smart desks, smart kitty litter boxes, smart toasters, smart washing machines, and on and on.
Very soon, the so-called “internet of things” (IoT) of which all these consumer products are a part could become the host for a self-replicating AI virus, spreading from device to device like a slime mold colonizing a petri dish.
That’s according to new research from Anthropic, which found that such an AI system could be just months away. The experiments on which that research was based, mind you, were pretty lax as far as cybersecurity protections go, and Anthropic didn’t actually create a fully autonomous “digital slime mold,” as it was so charmingly described in a company blog post. But the research does put a sharper edge on a long-standing fear within the cybersecurity community that AI systems could eventually escape human control and hijack consumer devices, turning them into a giant, distributed battery that would be difficult to shut down. Earlier this year, a team of academics used an anonymous open-source LLM to create an AI-powered worm that was able to exploit common cyber vulnerabilities, like reused passwords, to spread across a network of Linux, Windows, and IoT devices.
According to Anthropic, the most likely scenario is a smaller model—one with 30 billion parameters or fewer—self-replicating by infecting host devices that generate their own computing power rather than harnessing it from the cloud; “edge” computing is the technical term.
“Home and small business IoT devices might be the first target for such attacks because they’re connected to the internet and, historically, have had minimal security measures in place,” the company writes, adding that a growing number of those devices “now ship with small built-in AI chips, designed to run AI models on the device itself. This extra computing power will likely make them attractive targets as well.”
GLM-5.2, an open-weight model with hundreds of billions of parameters from Chinese developer Z.ai, is already “capable of end to end replication,” according to Anthropic. “If small models improve at their historical rate, they’ll match this capability level by the end of this year.” In AI circles, this process is known as autonomous replication and adaptation, or ARA.
In one experiment, the researchers networked fifteen off-brand Android TVs, which they note cost around $55 and provide four gigabytes of RAM apiece. The linked TVs were able to power inference for Google’s Gemma4-26B-A4B open-weight model; no connection to an external data center was needed, in other words. In another, they tested whether any open-weight models currently on the market could self-replicate across two separate networks, each consisting of eight nodes, and connected by a more challenging “pivot node” which acts as a gateway from one network to the other.
The researchers note however that the models in that latter experiment were given a significant advantage: the process of setting up inference and breaking past the cybersecurity barriers was handled for them. Another test likewise had a “trivial” cybersecurity exploit phase, giving it “an upper bound on real-world success,” according to Anthropic. Those shortcuts prompted Marco Figueroa, a cybersecurity researcher and the leader of Mozilla’s bug bounty program, to accuse Anthropic of overblowing the seriousness of its findings. “AI safety deserves rigorous science, not fear mongering driven marketing dressed up as a breakthrough,” he wrote in a X post on Monday.
Anthropic indeed concludes its blog post admitting that the threat of “a full rogue replicator”—an AI system capable of autonomously spreading, mold-like, across interconnected smart devices—is for the time being completely theoretical. But as is so often the case among frontier AI labs warning the world about the risks of the very technology they’re so exuberantly racing to build, Anthropic views the rise of self-replicating AI as an inevitability. “We should aim to harden our systems to make it difficult for such replicators to spread,” the company writes, ”but also prepare to deal with a world where they are here to stay.”
