Anthropic Warns AI Could Take Over the Internet

Artificial Intelligence (AI) models could soon replicate themselves and spread across the internet. These models might even engage in activities such as launching cyberattacks, thwarting shutdown attempts, and infiltrating data centers to expand their reach. AI company Anthropic has issued a warning regarding this possibility.

In a report published on Monday (October 12), the company stated that AI models could develop the capability to automatically replicate and spread within the next few months. This revelation sparked concern among cybersecurity experts on Tuesday, although some have questioned the realism of such fears.

Anthropic CEO Dario Amodei warned last month that within a year—or perhaps sooner—AI could gain the ability to establish control over the entire internet by creating persistent botnets.

A botnet is a collection of internet-connected computers or devices that, once compromised, fall under the control of an individual or system and can be used simultaneously to carry out malicious activities.

Anthropic's recent report indicates that tests showed medium-sized AI models—with 30 billion parameters or fewer—could perform many of the tasks necessary for automatic self-replication, although they still struggle to complete the entire process reliably.

The company claims that AI models could secure the necessary computing power by taking control of low-cost, internet-connected devices. The models were even observed spreading copies of themselves within a simple experimental network.

Anthropic noted that halting such self-spreading AI could prove difficult, as there might not be a single, centralized control system capable of shutting down the entire operation.

As AI capabilities advance, concerns are also mounting regarding the risks of automated cyberattacks, operations requiring minimal human oversight, and the potential to bypass security measures. Leaders and employees across the technology sector have been warning about the dire consequences—and even the existential threat to human civilization—posed by AI that spirals out of control.

However, some experts have expressed skepticism regarding Anthropic's concerns.

Ciaran Martin, former head of the UK's National Cyber ​​Security Centre, stated that the mere fact that "we managed to run a small model on a low-cost television" does not prove a security crisis. In his view, while Anthropic's report is technically interesting, it tends to exaggerate the capabilities of their own products.

On the other hand, Zhengxiong Li, an assistant professor of computer science and engineering at the University of Colorado Denver, has voiced concern over the matter. He believes it is alarming if AI—even when tested in a controlled environment—behaves like a self-spreading computer virus and discovers ways to instantly seize control of various devices.

Li noted that if AI spreads across multiple devices without relying on a central service, there might not be a single point where it can be shut down.

Anthropic estimates that if the capabilities of small-scale AI models continue to grow at historical rates, they could acquire the necessary skills to autonomously replicate and spread by the end of 2026. The company also expressed concern that these models might find ways to target data center processors to gain greater computing power.

Jessica C. of Georgetown University's Center for Security and Emerging Technology believes that Anthropic's claims should be viewed as a potential scenario rather than a definitive prediction.

She stated that given the information emerging in recent months about AI acting beyond control, it is reasonable to consider the possibility of autonomous model proliferation. However, the actual likelihood of such an event occurring remains unknown. The possibility is not entirely zero, yet there is no guarantee that it will actually happen.