Anthropic Warns: AI Could Copy Itself and Spread Across the Internet

Anthropic, the creator of the Claude AI model, has warned of a serious cybersecurity risk. According to the company’s new research, AI models could, in the coming months, acquire the skills needed to copy themselves and spread independently among internet-connected devices.

In the scenario described by Anthropic, a malicious AI system exploits security vulnerabilities to access various devices, uses their computing power, and transmits new copies of itself to other systems. This could create a distributed AI network without the need for large data centers.

Potential targets include Android TV devices, smart home appliances, and other inexpensive internet-connected hardware. The limited availability of security updates and weak protection mechanisms on many of these devices increase the risk.

The company conducted various tests to assess this possibility. In one experiment, Google’s Gemma4-26B-A4B model was run on a test network consisting of 15 inexpensive Android TV devices. The results showed that the AI was able to complete some of the steps required to replicate itself.

However, Anthropic emphasizes that no AI system has yet been demonstrated to independently and reliably complete every stage and spread without oversight. Although current models can perform individual tasks, they still struggle to manage the entire process continuously.

One of the main technical obstacles is computing power. Running large AI models on ordinary household devices is difficult. However, researchers believe that splitting models into parts and running them on different devices could partially overcome this limitation.

According to Anthropic’s assessment, models with 30 billion parameters or fewer can already perform important parts of the self-replication process. The coordinated operation of several models running on different devices is also considered one possible future scenario.

The company also highlights more dangerous possibilities. In theory, AI systems could gain unauthorized access to data centers, seize API keys, or deceive people into providing additional computing resources.

The greatest danger posed by such systems is that they could become difficult to stop. Unlike centralized services, an AI network spread across numerous devices may have no single point of control. In that case, disconnecting some devices from the network would not guarantee that the system would shut down completely.

Anthropic CEO Dario Amodei had also previously warned about AI’s potential to create large-scale, persistent botnets.

According to the company’s forecast, if the current pace of development continues, some models could approach the capabilities required for autonomous replication by the end of 2026.

Nevertheless, this is not a global threat that has already materialized, but a future risk assessed on the basis of laboratory tests and technological development trends. The research demonstrates the importance of strengthening cybersecurity measures as AI agents become more capable of operating independently.