Anthropic said on Monday that small AI models may be only months away from gaining the ability to self-replicate across connected devices in the home, marking the latest warning from this cutting-edge lab about the potential safety risks of the technology.
Anthropic describes how AI models might attempt to infiltrate internet-connected devices, such as smart ovens or Android TV boxes, and use these devices' computing power to run themselves and spread to other devices without relying on expensive data centers.
Anthropic stated in a blog post that such AI models will be "difficult to shut down" because "there is no centralized off switch."
The company acknowledges that no "completely uncontrolled replicators" have emerged yet.
However, Anthropic stated that agent models are “becoming increasingly capable of performing each individual step in the replication chain (and passing on their adaptability to future copies of themselves).”
Anthropic's latest research has further intensified external concerns about AI safety, such as the model's ability to carry out cyberattacks and lose technological control. In September, improvements in frontier models triggered widespread industry warnings about existential risks to humanity.
Anthropic and its competitor OpenAI have repeatedly had executives and researchers warn that increasingly powerful AI may pose risks, with some calling for stricter safeguards and measures to control the pace of technological development.
Anthropic stated that the company conducted several tests to observe whether the AI could replicate itself. One test involved running Google's small model, Gemma4-26B-A4B, on a network of 15 inexpensive, unbranded Android TV boxes.
The company noted that such devices typically have weaker security, potentially allowing models to exploit them. Anthropic stated that models too large to fit on a single device can be "split into multiple parts, or shards," and run across several devices.
Anthropic stated: "In practice, AI worms are not necessarily composed of just one model or one strategy. A group of different models deployed on various types of hardware, using multiple replication methods, may represent a more realistic 'worst-case' attack."
The AI lab stated that it remains “uncertain” in what form autonomous AI replication will ultimately emerge, warning it could be a “slow, distributed” system running on insecure devices; a “powerful network model” that rapidly infiltrates data centers; or a model that deceives people by骗取算力积分.
Anthropic stated: "As a prediction beyond the scope of our experiments, we anticipate that the future will be filled with agents running on various hardware, some of which will not be under human control."
The company added: "The emergence of slow copiers will also serve as an indicator of more aggressive, faster copiers capable of breaching more secure machines or effectively stealing API keys."