Anthropic Warns Small AI Models Could Be Months Away From Self-Replicating Across Smart Devices

According to Anthropic, small AI models could be just months away from being able to copy themselves onto inexpensive internet-connected devices, such as Android TV boxes, without relying on a data center.

The idea of AI models infecting connected devices is not new and has been discussed in online forums for some time. Now, in a study published on October 12, 2026, Anthropic shows through controlled experiments how this scenario could become possible.

How Anthropic Ran an AI Model Across 15 Android TV Boxes

In the first experiment, the company used distributed inference, a technique that splits a model’s execution across multiple devices. Each device processes part of the model’s layers and communicates with the others over a network. Using this approach, the researchers ran Gemma4-26B-A4B on 15 Android TV boxes, each costing approximately $55 and equipped with 4 GB of RAM.

The 26-billion-parameter model ran at a very limited speed because of the hardware: just 0.39 tokens per second. A token is a unit of text that can represent a word or part of one. At that speed, generating each token took approximately 2.5 seconds. Even so, the experiment showed that devices unable to run a model of that size individually could work together to run the complete model.

Testing AI Self-Replication Across Android Devices

In the second experiment, Anthropic investigated whether an advanced model could identify new devices, access them, install the necessary components, and continue replicating without human intervention. The test involved eight machines running emulated Android environments. The GLM-5.2 model reached all eight devices in four out of five attempts. In the fifth, it reached seven.

In this experiment, GLM-5.2 was tasked with installing Llama3.1-8B, a smaller model, on each machine rather than installing itself. Anthropic’s conclusion assumes that a small model with capabilities comparable to GLM-5.2 could perform the same process.

The company acknowledges that the test conditions were simplified compared with the real world. The devices had an ADB (Android Debug Bridge, an interface that allows Android devices to be controlled through commands) port open without authentication, making access easier. The model coordinating the process also ran on a separate server. The study focused on installing and running the models rather than exploiting vulnerabilities to break into the devices.

For these reasons, Anthropic considers the result an upper bound on the success rate that could be expected under real-world conditions. The company also acknowledges that it has not yet demonstrated a complete replicator capable of spreading autonomously across a real network.

However, Anthropic points out that older botnets (networks of infected devices), such as Mirai, spread using equally simple methods, including attempts to log in with factory-default passwords. In practice, the security of internet-connected devices may not make gaining access much harder than in the simplified experiment.

With those limitations in mind, the company estimates that small models could reach GLM-5.2’s demonstrated capabilities by the end of 2026 if they continue improving at their historical rate. Anthropic emphasizes that this is only a rough estimate.

Could AI Models Replicate Outside a Laboratory?

According to experts, the challenges would be considerable. In a real network, models would encounter unstable connections, different hardware architectures and operating systems, and the need to transfer large amounts of data between devices. That traffic could also attract the attention of security systems. Some experts question the scale of the risk, although they do not rule out the possibility of such systems emerging.

In an interview with Reuters, Ciaran Martin, former head of the UK’s National Cyber Security Centre, questioned Anthropic’s warning:

“’We managed to run a small model on cheap TVs,’ does not a security crisis make.”

Other experts expressed greater concern. Zhengxiong Li, an assistant professor of computer science and engineering at the University of Colorado Denver, told Reuters:

“If it runs across many devices without depending on a central service, there may be no single place to shut it down.”

Jessica Ji, a senior research analyst at Georgetown University’s Center for Security and Emerging Technology, considers it reasonable to investigate this possibility given recent advances in AI. However, she emphasizes that the likelihood of autonomous replication actually happening remains unknown.

Based on the study and the reporting surrounding it, the results appear to serve more as a warning about a possible scenario than a prediction of what will necessarily happen. Reuters presents a similar assessment: an unnamed researcher said it made sense to view Anthropic’s publication as more of a hypothetical than a prediction.

How Cybersecurity Companies Could Prevent AI Replication

For cybersecurity companies, this is an opportunity to start developing protections for devices that have historically received little attention when it comes to security.

The study itself suggests several approaches. Traffic generated by a self-replicating system could leave traces detectable by internet service providers. Quarantining infected devices has already proved effective against Mirai, while measures such as the UK’s ban on default passwords for connected devices could reduce the number of vulnerable targets.

It remains unknown whether AI models will be able to spread autonomously across the internet, or when that might happen. Anthropic’s experiments do show, however, that some of the capabilities required already exist separately.