Anthropic On Monday, it was indicated that small AI models may only have a few months before gaining the ability to self-replicate among connected devices at home. This is also the latest warning issued by this cutting-edge laboratory regarding the potential security risks of this technology.

Anthropic describes that the AI model may attempt to infiltrate connected devices, such as smart ovens or Android TV boxes, and utilize the computing power of these devices to operate itself and spread to other devices, without relying on expensive data centers.

In a blog post, Anthropic stated that such AI models will be “very difficult to shut down,” “because there is no centralized point of shutdown.”

The company acknowledges that so far, there have been no “completely out-of-control replicators.”

However, Anthropic indicates that proxy models are 'becoming increasingly capable of completing each individual step in the replication chain (and passing on their adaptability to future copies of themselves).'

The latest research on Anthropic has further exacerbated external concerns regarding the security of AI, such as the model's ability to carry out cyberattacks and the potential for technological runaway. In September, improvements to cutting-edge models triggered warnings across the industry about the risks to human survival.

Executives and researchers from Anthropic and its competitor OpenAI have repeatedly warned that the increasingly powerful AI could pose risks. Some of them have called for the establishment of stricter safeguards and measures to control the pace of technological development.

Anthropic indicates that the company conducted multiple tests to observe whether AI is capable of self-replication. One of these tests involved running a small model of Google, Gemma4-26B-A4B, on a network composed of 15 inexpensive, non-branded Android TV boxes.

The company pointed out that such devices usually have weaker security, which may provide an opportunity for models to penetrate them. Anthropic indicates that models that are too large to fit into a single device can be “divided into several parts, or fragments (shards),” and run on multiple devices.

Anthropic indicates that: 'In practice, the AI worm is not necessarily composed of just one model or one strategy. A group of different models deployed on various types of hardware, using multiple replication schemes, may represent a more realistic 'worst-case' attack.'

This AI laboratory stated that it is still “uncertain” in what form autonomous AI replication will ultimately take, and warned that it could be a “slow, distributed” system running on insecure devices; it could also be a “powerful neural network model” that invades data centers at a rapid speed; or it could be a model that deceives people by exploiting computing power points.

Anthropic indicates: 'As a prediction that goes beyond the scope of our experiments, we anticipate that the future will be filled with agents running on various types of hardware, some of which will not be under human control.'

The company added, "The emergence of slow replicators will also serve as a barometer for more aggressive and faster replicators, which are capable of breaking into machines with stronger protections or effectively stealing the API keys."