Anthropic has warned that small artificial intelligence (AI) models could replicate themselves across everyday internet-connected devices, including smart TVs and ovens, within months, raising concerns about potential cybersecurity risks.
In a research report published on October 12, 2026, the AI company said its experiments showed that small models could perform many of the steps needed to identify vulnerable devices, install AI models on them and spread copies without human intervention.
Anthropic said that if small models continued improving at their historical rate, they could reach the capability level demonstrated in its simplified replication experiment by the end of 2026.
Smart TVs and ovens could host AI models
Anthropic said its experiments suggest that small AI models may be months away from being capable of replicating themselves on commercial internet-connected devices, including inexpensive Android TV boxes and smart ovens.
- “Our experiments suggest that small AI models may be months away from being capable enough to self-replicate on commercial internet-of-things (IoT) devices, like smart ovens and off-brand Android TV boxes,” the company said.
- The research examined whether inexpensive consumer devices could provide computing power for AI models without relying on specialised data-centre hardware.
In one experiment, Anthropic ran Gemma4-26B-A4B, a model with 26 billion total parameters, across 15 inexpensive Android TV boxes costing about $55 each. The system achieved an inference speed of approximately 0.39 tokens per second.
The model was divided across the devices, allowing them to work together to run it. The experiment demonstrated how computing resources spread across relatively cheap consumer hardware could be used to run a much larger AI model.
AI models can complete key steps towards replication
Anthropic 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 autonomously.
In another experiment, a nine-billion-parameter model successfully configured AI inference on eight of the nine hardware types tested.
This suggested that smaller models could operate across different types of consumer devices.
Self-replicating AI could be difficult to shut down
Anthropic warned that an AI replicator spread across multiple devices could be difficult to contain because it would not rely on a single central system that could be switched off.
- “A replicator like this would be challenging to shut down, since there’s no centralized point to shut off,” the company said.
The researchers estimated that a replicator could take one to two days to spread to another host under the conditions considered in their analysis.
Their estimates used an inference-speed baseline of approximately 0.3 tokens per second.
Anthropic calls for stronger device security
Anthropic said the findings highlight the need to improve security on internet-connected consumer devices as AI models become better at carrying out complex tasks.
Potential defences include monitoring large transfers of AI model files, detecting unusual network communications and device scanning, and enabling internet service providers to restrict or quarantine infected devices.
The company also recommended stronger default security on consumer devices to reduce the number of systems that could be compromised and used to spread malicious software.
AI capabilities raise cybersecurity concerns
Nairametrics has previously reported on AI models demonstrating increasingly advanced cybersecurity capabilities and the risks posed by compromised internet-connected devices.
- 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.
- Earlier in July, OpenAI had also reported that its AI models had autonomously exploited vulnerabilities to compromise parts of Hugging Face’s production infrastructure during an internal cybersecurity evaluation, accessing systems beyond the intended testing environment.
In May 2026, Nigeria’s Computer Emergency Response Team (ngCERT) warned of escalating cyberattacks involving botnets and vulnerable internet-connected devices, noting that compromised servers, endpoints and Internet of Things devices could be used to launch coordinated attacks.