The risk of AI self-replication has come under scrutiny following a study by Anthropic

On October 12, researchers from Anthropic, the company behind the Claude model, published a study warning of the dangers posed by artificial intelligence systems that can copy themselves and spread across the internet without human intervention. The experiments show that even relatively small models can already perform several steps necessary for this process, indicating a growing challenge for cybersecurity.

What does it mean for an AI to replicate itself?

AI self-replication occurs when a system can transfer its components to other computers, install what it needs to function, and repeat the process on new machines autonomously. The concept is similar to that of a worm, a type of malware that spreads between connected devices, but with the added ability to adapt its actions to different environments, identify available tools, and solve technical problems during propagation.

What did Anthropic discover in its tests?

In a study published on its research website, Anthropic tested whether smaller models could perform steps of autonomous replication without relying on large data centers. The researchers assessed the ability to configure models on unknown equipment and complete a replication chain.

According to the company, models with up to 30 billion parameters can already perform several individual tasks necessary for the process, although they still make mistakes when combining all the steps. In simplified tests, the GLM-5.2 model from Chinese company Z.ai and more advanced models were able to complete the end-to-end cycle.

The experiments also showed that a 26 billion-parameter model could be distributed across 15 TV box devices—inexpensive devices connected to a television—to investigate whether they could provide sufficient computing power to run AI models.

It is important to note that the tests do not prove the existence of an AI that spreads freely across the internet. They assess technical capabilities under specific experimental conditions, without demonstrating that a system has spontaneously compromised millions of real devices.

Why are TVs, cameras, and appliances part of this discussion?

The warning does not apply only to computers and corporate servers. Internet-connected devices such as security cameras, televisions, vacuum robots, and smart appliances can also have exploitable computing resources.

Many of these devices receive less security attention than computers and mobile phones. Weak passwords, infrequent updates, and software flaws can make them vulnerable to hacking.

A network of compromised devices can form a botnet—a collection of remotely controlled devices used to perform coordinated activities, such as attacks on websites or other digital infrastructure. With AI, this type of threat could gain the ability to adapt, trying to understand different configurations and adjust its approach to continue operating.

For professionals and companies, this highlights the importance of updating equipment, changing default passwords, restricting access, and monitoring unusual behavior on networks.

How can professionals prepare for the risks of AI?

The discussion about self-replication shows that understanding artificial intelligence involves more than learning to write commands or automate tasks. It also requires understanding the limits, permissions, and risks associated with using these tools.

For those working in technology, information security, or management, this includes assessing which systems can access files, execute code, and interact with external services. In other areas, it is important to know the company’s AI usage policies and avoid granting unnecessary permissions to digital tools.

Can AI already replicate itself without human control?

Anthropic believes that the evolution of models could make autonomous replication a more concrete possibility in the near future. In the study, the company estimates that smaller models could reach the level of capability observed in the tests by the end of 2026, if they maintain the historical pace of evolution. However, this prediction is not a guarantee that the scenario will materialize.

The main concern is that a decentralized network of compromised devices would be more difficult to disable than a system concentrated on a single server. At the same time, the feasibility of real attacks depends on factors such as access to equipment, exploitable vulnerabilities, computing resources, and the ability to maintain operation.

The situation still requires investigation and preventive measures. Monitoring the evolution of artificial intelligence and understanding its technical risks will become increasingly important for professionals and organizations that rely on connected systems.