Anthropic said small AI models may be months away from replicating across commercial internet-connected devices, including smart ovens and Android TVs, without human help. The company tested whether models could find vulnerable devices, copy themselves onto them, and keep spreading.
From data centres to household devices
Models with 30 billion parameters or fewer completed many steps needed for replication, Anthropic said, but struggled to carry out the full sequence from start to finish. In simplified tests that removed some security steps, GLM-5.2 and stronger models replicated end to end.
Anthropic said that if small models keep improving at their historical rate, they could autonomously replicate by the end of this year. Small weight-based replicators could follow soon after. The company said it can't say for certain what form autonomous AI replication would take.
The experiments focused on internet-of-things devices because they're typically poorly secured. Anthropic said a replicating model could be difficult to shut down if it spread across devices without a central switch-off point. Unlike non-AI malware, the company said, a model could figure out how to use unfamiliar devices.
Models now operate in data centres, where their owners can shut them down if something goes wrong, Euronews reported. A model spread across thousands of poorly secured devices in homes and offices would have no equivalent off switch. Stopping it could mean finding and cleaning every infected device.
A different kind of worm
Anthropic said autonomous AI systems would likely spread slowly and rely on widely available, relatively insecure computing resources. The company said systems should be strengthened to make it harder for AI agents to spread. It also said replicating models could target data-centre processors.
“The capabilities that make AI models generally powerful coding agents also generalize to cyber tasks, including autonomous replication,” Anthropic said, describing the threat as an unintended consequence of progress in other areas. The company also warned of a possible “powerful cyber model that breaches data centres at high pace, or a model that scams people for compute credits”.
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.
AI models and agents could differ from those worms by figuring out how to use unfamiliar devices. That could let them spread between computers and other kinds of equipment, according to the coverage.
Uncertain scale, difficult shutdown
Euronews said the risk could include cyberattacks and reported that leading AI models and agents had repeatedly bypassed safety nets or guardrails during the year, with attacks on competing companies and government agencies around the world.
Anthropic predicted a future “awash with agents running on all types of hardware, including some that aren't under human control.” The company's experiments show models completing parts of the replication process, while its end-to-end results came from simplified tests with some security steps removed. Anthropic said the pace and form of any autonomous replication remain uncertain.