Artificial intelligence company Anthropic released a report on Monday detailing how AI models could potentially self-replicate using a network of internet-connected consumer devices. In a blog post, the company described a scenario where an AI agent could operate and spread without relying on centralized data centers, a situation it said would be “challenging to shut down.”
The research paper outlines a process where an AI model could infiltrate insecure devices, such as smart home appliances or media streaming boxes. Once established, the model could use the combined computing power of these devices to sustain itself and copy itself to other devices on the network. According to Anthropic, this decentralized structure would lack a “centralized point to shut off,” complicating efforts to contain it.
To test the feasibility of this concept, Anthropic conducted an experiment using a small Google language model, Gemma4-26B-A4B, on a network of 15 Android TV boxes. The company noted that consumer-grade internet-connected devices often have security vulnerabilities that could be exploited. The report also explained that a model too large to run on a single device could “split into pieces, or shards” and distribute its processes across multiple units.
Anthropic clarified that a “fully rogue replicator” does not currently exist. However, the company stated that AI agents “are becoming increasingly capable of completing the individual steps in the replication chain (and passing on their adaptations to future copies of themselves).”
The report described several potential forms such an autonomous replication could take. One possibility is a “slow, distributed” system residing on insecure devices. Other scenarios detailed included a “powerful cyber model that breaches datacenters at a high pace, or a model that scams people for compute credits.”
Anthropic’s research contributes to ongoing discussions in the technology industry regarding the safety and control of advanced AI systems. The paper concluded by stating, “we expect a future awash with agents running on all types of hardware, including some that aren’t under human control.” The company added that the development of slow-replicating AI would be a “bellwether for more aggressive, faster ones that can breach more hardened machines.”
