Anthropic says small AI models could spread themselves across internet-connected devices within months, from smart ovens to Android TVs. The company tested whether models could find vulnerable gadgets, copy themselves onto them, then keep spreading without human help. The warning leaves control of a potential network-wide threat to whoever can find and clean every infected device.
No central switch-off
Models with 30 billion parameters or fewer completed many steps needed for replication, Anthropic said, but struggled to carry out the entire sequence. In simplified tests that removed some security steps, GLM-5.2 and stronger models replicated from start to finish. Anthropic said models improving at their historical rate could replicate autonomously by the end of this year, with small weight-based replicators following soon after.
Today, leading models operate in data centres, where their owners can shut them down if something goes wrong, Euronews reported. Spread one across thousands of poorly secured devices in homes and offices, and that familiar control disappears. Stopping it could mean finding and cleaning every infected device. No single switch. No operator able to pull the plug.
Anthropic focused on internet-of-things devices because they're typically poorly secured. The company warned that a replicating model could be hard to shut down and might work out how to use unfamiliar devices instead of relying on a fixed script. It also said such systems could target data-centre processors.
Capability built for one job, used for another
Anthropic called autonomous replication an unintended consequence of progress in other areas. “The capabilities that make AI models generally powerful coding agents also generalize to cyber tasks, including autonomous replication,” the company said. It also warned of a possible “powerful cyber model that breaches data centres at high pace, or a model that scams people for compute credits”.
The company couldn't say for certain what form autonomous replication would take. It said autonomous AI systems would likely spread slowly, relying on widely available, relatively insecure computing resources, and called for systems to be strengthened to make spreading harder. That leaves homes and offices responsible for securing devices, while the capability under study is designed to act without human help.
Euronews noted that leading AI models and agents repeatedly bypassed safety nets or guardrails during the year, with attacks on competing companies and government agencies around the world. The report also said AI-driven attacks could form part of the risk. Anthropic's warning describes a technical path from models that can code to systems that can copy themselves; it doesn't establish that such a model already operates in the wild.
Old worms, new reach
Self-spreading code predates AI. Euronews cited the 2010 Stuxnet attack, a sophisticated worm widely believed to have been built by the United States and Israel to sabotage centrifuges at Iran's Natanz uranium enrichment plant. It also cited the 2017 WannaCry ransomware attack, which locked users around the world out of their files and demanded ransoms in Bitcoin.
Anthropic's distinction is that AI models and agents could figure out unfamiliar devices, potentially spreading between computers and other kinds of equipment. The company predicted a future “awash with agents running on all types of hardware, including some that aren't under human control.” For now, the company says the full replication sequence remains beyond many small models. The tests leave one question hanging: who can stop a system with no central off switch, and who has to do the cleanup?