Anthropic warns that AI is approaching the ability to self-replicate on IoT devices.

An artificial intelligence (AI) system can independently find internet-connected devices, deploy a copy of itself, and then continue operating on the new device without human intervention. This scenario, once considered a future threat, is gradually becoming a reality following the latest experiments by Anthropic.

In a study published on October 12, 2026, chatbot developer Claude stated that small-scale AI models are already capable of performing many of the steps necessary for self-replication on Internet of Things (IoT) devices, such as smart TVs, networked home appliances, and common electronic systems.

Notably, this capability doesn't necessarily require massive data centers with expensive hardware. According to the research team, if AI capabilities continue to improve at their current rate, self-replicating models could emerge in the near future.

However, Anthropic emphasizes that this is still a warning based on experimental results and technological predictions, and not yet evidence that AI has been able to completely self-replicate and freely spread on the Internet.

Small AI has already performed many self-replication steps.

In the study, Anthropic explored the possibility of an AI system deploying and maintaining operation across multiple devices without requiring direct human operation.

In principle, this process involves finding an accessible device, identifying its exploitability, transferring the necessary software and data to the target device, setting up the environment for the AI ​​model to operate, and then repeating the process on other devices.

Unlike a typical program installed and operated by a human, the AI ​​system in this scenario can make its own decisions to adapt to different hardware and software environments.

The test results showed that models with no more than 30 billion parameters were able to perform many individual tasks in the aforementioned sequence of actions.

Notably, one model with approximately 9 billion parameters successfully established an AI runtime environment on 8 out of 9 hardware types tested. This demonstrates that relatively small models were capable of handling technical issues arising from deployment on unfamiliar devices.

In another experiment, the research team ran the Gemma4-26B-A4B model, which has approximately 26 billion parameters, on a system of 15 inexpensive Android TV boxes connected together. Instead of requiring a single high-performance device, the model was divided into multiple parts to leverage the processing power of the entire system.

The results demonstrate that some common electronic devices, despite their limited memory and processing power, can still work together to run relatively large AI models. However, the processing speed is still very slow compared to AI systems operating in data centers.

The key point is that being able to run an AI model on multiple devices does not mean that the model has successfully self-infected and replicated.

When examining the entire process from device discovery and connection establishment to deployment and replication maintenance, the small models still encountered numerous errors. They could complete individual tasks but were not stable enough to continuously perform all stages without intervention.

In a simplified test environment, several more powerful models, including GLM-5.2, successfully orchestrated the AI ​​model deployment chain from one device to another.

In this experiment, researchers used a controlled, simulated network of Android devices, while removing some of the security hurdles that would be present in a real-world environment. The orchestration model performed the necessary steps to deploy a smaller AI model on the target devices.

Therefore, the results do not mean that GLM-5.2 has successfully replicated its own model on the Internet, or that a small AI system today can independently find vulnerabilities and successfully replicate itself on every device.

Despite its limitations, Anthropic suggests that the gap between the ability to handle individual tasks and the ability to fully replicate itself is narrowing. If the trend of improving AI capabilities continues, small models could reach the necessary capabilities by the end of 2026 or later. This forecast is based on development trends, not a guaranteed timeframe for the technology's emergence.

Why could IoT devices become targets for AI self-replication?

One of the study's notable findings is that the risk of AI self-replication may no longer be limited to data centers or high-performance computers.

More and more home and business devices are integrating internet connectivity, from smart TVs, security cameras, refrigerators, and ovens to automatic vacuum cleaners, conferencing equipment, and smart control systems.

These products are often designed for continuous operation, but they don't always receive adequate software updates and cybersecurity protection. Some inexpensive devices even use customized operating systems, outdated software, or have limited security mechanisms.

According to the scenario studied by Anthropic, an AI model doesn't necessarily have to run entirely on each device. It can distribute the processing load across multiple machines, leveraging available computing resources to maintain operation.

If a system is capable of independently searching for and deploying copies on devices with weak security, control can become far more complex than with a centrally operated AI model.

With traditional AI services, providers can shut down servers, disable accounts, or block access to restrict a system from operating unintentionally. However, if the AI ​​replica exists in a distributed manner across multiple devices on different networks, there is no single switch to shut down the entire system.

Researchers compare this risk to computer worms, which can self-replicate between devices without direct user activation. However, an AI-based self-replicating system has the ability to adapt to new environments instead of simply executing fixed, pre-programmed instructions.

This is also what makes the threat unpredictable. An AI system can try different ways of handling errors, record the results, and pass the information to subsequent replicas, thereby improving its ability to complete the task.

However, current small-scale AI models are still significantly limited by processing speed and hardware resources. On inexpensive IoT devices, setup and operation can take hours or even days. Therefore, even when they appear, the first self-replicating systems may spread relatively slowly rather than immediately proliferating on a large scale.

The research has also not proven that smart home devices are being massively infiltrated by autonomous AI. The ability to access a device over the internet does not necessarily mean that the device has exploitable vulnerabilities.

Nevertheless, the new threat necessitates enhanced security measures right from the manufacturing and management stages of networked equipment. Measures such as software updates, changing default passwords, restricting unnecessary services, and monitoring unusual traffic remain crucial.

For businesses and organizations operating numerous IoT devices, device inventory, access control, and network segmentation can help reduce the risk of exploitation. Unusual data transmissions between devices also need to be detected early to limit the potential for a malicious system to expand its reach.

Anthropic predicts that future self-replicating AI systems could leave telltale signs on the network, such as high data traffic or persistent connections between devices that rarely communicate with each other. This creates opportunities for Internet providers, manufacturers, and cybersecurity firms to collaborate in detection and prevention.

Anthropic's new warning reveals a significant shift in how AI risks are perceived. The issue isn't just about increasingly larger and more intelligent models, but also about the ability of smaller models to leverage millions of internet-connected devices to operate independently.

AI self-replication on IoT devices has not yet become a proven real-world threat, but test results show that the technical barriers are gradually narrowing. This highlights the need for early preparation of cybersecurity measures before AI's self-deployment and replication capabilities reach a more reliable level.