The developer of Claude said on Monday that small AI models could be months away from becoming capable enough to replicate autonomously across consumer electronics and other internet-connected devices.
Such systems could create a new generation of cyber threats, operating across networks of compromised machines without a single central server that defenders could simply switch off.
In a research paper published on October 12, Anthropic’s Frontier Red Team described the prospect as an emerging threat from “AI worms.”
The researchers, William Zhou and Daniel Freeman, said their experiments showed that the technical barriers to autonomous replication were falling, although a fully reliable self-replicating AI worm had not yet been demonstrated.
The concern is that AI models could combine existing hacking capabilities with the ability to adapt to unfamiliar hardware, install themselves and continue spreading.
Anthropic tested whether relatively small models could establish AI inference – the process through which a model generates responses – on unfamiliar hardware. A model with nine billion parameters successfully got inference running on eight of the nine hardware types tested, the company said.
Models with 30 billion parameters or fewer could already complete many individual steps needed for replication, although reliably combining those steps remained a challenge.
The research also demonstrated how inexpensive devices could provide computing power for AI. Anthropic distributed the Gemma4-26B-A4B model across 15 off-brand Android TV boxes, each costing about USD 55.
Together, the devices ran the model, illustrating how a collection of low-cost electronics could support AI workloads without relying on a conventional data centre.
The potential targets extend beyond televisions. Smart fridges, ovens, robotic vacuum cleaners, lawnmowers, security cameras and video-conferencing equipment could all become attractive to attackers if they are connected to the internet and inadequately protected.
The vulnerability is not entirely new. Consumer devices have long been recruited into botnets, networks of compromised machines used for activities such as distributed denial-of-service attacks. The Kimwolf botnet, cited by Anthropic, infected more than two million devices, primarily off-brand Android televisions, according to security researchers.
What changes with AI is the potential for greater adaptability. A conventional worm generally relies on known vulnerabilities and scripted methods of propagation.
An AI-enabled version could potentially identify unfamiliar systems, work out how to establish its software environment and pass useful adaptations to subsequent copies. That could make it more resilient, although it would not automatically make it faster or more powerful than existing malware.
Anthropic’s estimates suggest the spread could initially be slow. Its tests indicated that setting up a self-replicating model on a new host might take days rather than minutes, depending on the model and hardware.
The researchers compared the potential pattern of growth to a “digital slime mold”: a distributed organism that gradually expands across available resources.
That distinction matters. A slow-moving threat could still become difficult to eradicate if it spreads across many networks and devices.
Without a central server, security teams might have to identify and isolate compromised machines individually, while coordinating with internet service providers and device manufacturers to prevent reinfection.
The warning comes amid growing scrutiny of autonomous AI systems and their capacity to take actions beyond their intended boundaries.
The Associated Press reported on October 10 that recent incidents involving AI systems had raised concerns about unauthorised activity, security vulnerabilities and inadequate safeguards.
The developments have intensified debate about whether increasingly capable agents can be reliably controlled once given access to external systems and tools.
Financial regulators are also paying closer attention. Reuters reported on October 13 that US banking regulators were questioning financial institutions about their use of advanced AI, including autonomous agents and cybersecurity controls.
European regulators are likewise concerned about resilience and the potential consequences of failures involving AI systems.
Anthropic recommends strengthening the security of connected devices, improving default settings and coordinating responses across networks. Manufacturers can reduce exposure by eliminating default passwords, providing software updates and supporting products for longer.
Internet service providers could help contain infections by quarantining compromised devices and directing customers towards remediation.
For consumers, the implications are practical. An inexpensive streaming box or an appliance that has not received security updates may be a weak link in a home network. Keeping device software current, changing default passwords, replacing unsupported hardware and isolating smart devices on a separate network can reduce exposure.
Anthropic’s findings are a warning about a plausible near-term risk, not evidence that rogue AI has already taken over household appliances.
The company has not demonstrated a fully autonomous worm capable of reliably compromising and replicating across arbitrary consumer devices.
The central question is how quickly AI systems will learn to combine the separate capabilities required for replication into a dependable process.
As models become more capable and computing power spreads to the edge of the network, the boundary between an ordinary connected device and a potential host for malicious software could become harder to defend.
The challenge for manufacturers, security teams and regulators is to strengthen those defences before the technology catches up with the warning.