Smart TVs, security cameras and other connected household devices could become hosts for self-replicating AI worms capable of spreading without human intervention, Anthropic has warned.
In research published on 12 October 2026, the AI company demonstrated that powerful models can run across inexpensive Android TV boxes, raising concerns about future malware that could operate without centralised data centres.
The findings suggest smaller AI models are approaching some of the capabilities needed to replicate across vulnerable devices. Anthropic estimates that, if small models continue improving at their historical rate, they could reach the capability level required for autonomous replication by the end of 2026.
However, the threat remains largely hypothetical. Researchers have not demonstrated a fully self-contained AI worm autonomously spreading through ordinary household networks. Anthropic also acknowledges that the models currently capable enough to complete its simplified replication test are too large to run on the inexpensive devices examined in the research.
How 15 Cheap TV Boxes Ran a Powerful AI Model
Anthropic researchers William Zhou and Daniel Freeman investigated whether AI models could operate without expensive data centres. Their experiments examined whether ordinary connected devices could collectively provide enough computing power to run capable AI systems.
The team connected 15 off-brand Android TV boxes costing approximately $55 each. Together, the devices ran Google's Gemma4-26B-A4B model at 0.39 tokens per second. The model was distributed across the network, with individual boxes holding only part of it.
Although considerably slower than a typical chatbot, the experiment demonstrated that a substantial AI model could operate across relatively inexpensive hardware without specialised data-centre processors.
The researchers separately tested whether smaller models could configure AI software on unfamiliar machines. Qwen3.5-9B successfully established AI inference on eight of nine hardware types tested.
Inference involves running a trained model to generate responses or perform tasks. The result suggests smaller AI systems are becoming better at adapting to different computing environments.
Could AI Worms Spread Without Anyone Controlling Them?
Traditional computer worms exploit security weaknesses to copy themselves between machines. Anthropic examined a more adaptable possibility: AI systems capable of identifying targets, establishing computing environments and continuing to replicate with limited human involvement.
In a separate experiment, researchers used a firewalled network of eight Google Cloud nodes running Android emulators. GLM-5.2 reached all eight nodes in four of five trials and seven in the remaining trial while orchestrating a simplified replication process.
However, the experiment had significant limitations. The simulated devices exposed an unauthenticated Android Debug Bridge port, deliberately making access easier. Anthropic said this made the result an upper bound on likely real-world success.
GLM-5.2 also did not copy and run itself on those machines. The larger model operated through a separate inference server while the target nodes were configured to run the smaller Llama3.1-8B model. The evaluation system then allowed the replication process to continue after checking that inference had been successfully established.
The results therefore do not demonstrate a self-contained AI worm infecting ordinary smart TVs. Anthropic instead argues that future smaller models could eventually combine the required capabilities. Its end-of-2026 timeline is a rough projection based on historical improvements in general model capability, which the researchers use only as a proxy for replication ability.
Why Smart TVs and Household Appliances Could Become Targets
The potential targets extend beyond televisions. Anthropic identified devices such as smart ovens and other internet-connected equipment as possible sources of computing power for future AI worms. It also discussed cameras and other edge devices equipped with processors capable of running AI workloads.
Many connected devices contain processors capable of running software locally. Poorly secured products, particularly those with weak authentication or inadequate security support, could offer attractive targets for conventional malware as well as more advanced threats.
Anthropic cited the Kimwolf botnet as an example. The botnet reportedly infected more than two million devices, primarily inexpensive Android TV boxes. Kimwolf was conventional malware rather than a self-replicating AI system, but its reported scale illustrates how widely malicious software can spread across poorly secured consumer electronics.
A future AI worm could potentially distribute its computing workload across large numbers of compromised devices. Yet Anthropic's estimates suggest such replication could be slow. Its research calculated that establishing another working 9-billion-parameter model on inexpensive hardware could take more than one or two days.
Cybersecurity Experts Divided Over Anthropic's Warning
Not every cybersecurity expert believes the findings indicate an imminent crisis. Ciaran Martin, former head of Britain's National Cyber Security Centre, questioned the significance of the warning. Speaking to Reuters, he described the work as technically interesting while suggesting Anthropic's interpretation contained an element of vendor exaggeration.
Others expressed greater concern. Zhengxiong Li, an assistant professor of computer science and engineering at the University of Colorado Denver, told Reuters that an AI able to operate across many devices without depending on a central service could become difficult to shut down.
Jessica Ji, a senior research analyst at Georgetown University's Center for Security and Emerging Technology, also cautioned against treating autonomous AI proliferation as inevitable. She acknowledged the possibility but stressed that its likelihood remains uncertain.
Their differing assessments highlight an important distinction. Demonstrating that AI can perform replication tasks under deliberately simplified experimental conditions does not establish that it can reliably compromise and spread through protected consumer devices.
What Smart Device Owners Should Know
Anthropic's findings do not mean existing smart TVs have suddenly become infected with self-replicating AI worms. The research identifies a potential emerging cybersecurity threat rather than a confirmed outbreak.
The UK already requires manufacturers of many internet-connected consumer products to meet minimum security standards. Regulations that took effect on 29 April 2024 require passwords to be unique to each product or defined by the user. Manufacturers must also provide information on reporting security vulnerabilities and disclose how long security updates will be supplied.
Consumers can reduce their exposure by installing security updates, replacing unsupported devices and being cautious with inexpensive connected products from manufacturers with unclear security practices.
Anthropic's experiments demonstrate several components that could eventually support autonomous AI replication, but not a complete self-contained worm spreading through household networks. For now, the crucial gap remains between models small enough to run on inexpensive connected devices and models capable enough to reliably complete the entire replication process.
