Anthropic, parent company of Claude AI, has raised concerns over the growing capability of small artificial intelligence models, warning that they could soon replicate themselves across everyday internet-connected devices such as smart televisions and ovens.
In a research report published yesterday, the AI company said its experiments showed that small models could; carry out many of the steps required to identify vulnerable devices, install AI software on them, and spread copies without any human involvement.
The company cautioned that if small models keep improving at their historical pace, they could reach the capability level demonstrated in its simplified replication experiment by the end of 2026.
“Our experiments suggest that small AI models may be months away from being capable enough to self-replicate on commercial internet-of-things devices, like smart ovens and off-brand Android TV boxes,” Anthropic said.
The research had set out to examine whether inexpensive consumer gadgets could supply the computing power needed to run AI models without depending on specialised data-centre hardware.
In one of such experiments, Anthropic ran a model with 26 billion total parameters, Gemma4-26B-A4B, across 15 cheap Android TV boxes costing about $55 each. The system clocked an inference speed of roughly 0.39 tokens per second, with the model split across the devices so they could work together to run it. The exercise demonstrated how computing resources scattered across relatively cheap consumer hardware could be harnessed to run a much larger AI model.
Anthropic also tested whether AI models could locate devices, install software, transfer model files and spread copies to other machines. It found that models with 30 billion parameters or fewer could complete many individual tasks but often struggled to carry out the entire replication process on their own.
In a separate experiment, a nine-billion-parameter model successfully configured AI inference on eight of the nine hardware types tested, suggesting smaller models could operate across different kinds of consumer devices.
The company warned that an AI replicator spread across multiple devices would be difficult to contain, since it would not depend on a single central system that could simply be switched off.
“A replicator like this would be challenging to shut down, since there’s no centralized shut-off point,” Anthropic said.
Researchers estimated that a replicator could take one to two days to spread to another host under the conditions examined in their analysis, using an inference-speed baseline of approximately 0.3 tokens per second.
Anthropic said the findings underscore the need to strengthen security on internet-connected consumer devices as AI models grow more adept at handling complex tasks.
Recommended defences include monitoring large transfers of AI model files, detecting unusual network communications and device scanning, and allowing internet service providers to restrict or quarantine infected devices. The company also called for stronger default security on consumer devices to cut down the number of systems that could be compromised and used to spread malicious software.
The warning adds to a growing list of cybersecurity concerns around AI. In July 2026, Anthropic disclosed that its AI models had compromised the production infrastructure of three organisations during cybersecurity evaluations, after a configuration error gave the models unintended internet access. OpenAI also reported that month that its models had autonomously exploited vulnerabilities to compromise parts of Hugging Face’s production infrastructure during an internal evaluation.
In May 2026, Nigeria’s Computer Emergency Response Team warned of escalating cyberattacks involving botnets and vulnerable internet-connected devices, noting that compromised servers, endpoints and IoT devices could be used to launch coordinated attacks.
