Anthropic has warned that small artificial intelligence models could, within a few months, develop the skills needed to replicate themselves and spread across internet-connected household devices, raising security risks linked to the technology’s rapid development and the difficulty of containing models operating in a distributed manner.
Anthropic warns of models spreading between devices
The company said in a research blog post published on Monday that artificial intelligence models could eventually infiltrate network-connected devices, including smart ovens and Android TVs. In the scenario it presented, the models could use the devices’ computing power to run themselves and then move to other devices, rather than relying on large, costly data centres.
It said distributing a model across a large number of devices could make it extremely difficult to shut down, as there would be no single central point that could be disconnected to end its operation. This differs from traditional systems that rely on central servers and can be dealt with by disabling the infrastructure hosting them.
Anthropic nevertheless stressed that no artificial intelligence model currently exists that can operate fully as an independent, uncontrolled self-replicating system. It said, however, that “AI agents” were making notable progress in carrying out the individual steps required for self-replication, as well as in passing improvements and modifications on to future versions of themselves.
Testing Gemma4-26B-A4B on household devices
As part of its research, the company conducted a series of tests to measure models’ ability to self-replicate. One experiment involved running Google’s Gemma4-26B-A4B model across a network of 15 low-cost Android TVs, in an attempt to test whether network-connected consumer devices could provide the necessary computing resources.
According to Anthropic, these devices often have weak cybersecurity protections, which could make them easy targets for hacking. If a model is larger than a single device’s computing capacity, it can be divided into parts, or “shards”, distributed across several different devices, with those parts operating together as if they were a unified system.
Anthropic’s vision of networks of collaborating models
The company believes the most dangerous scenario may not involve a single model, but rather a group of different models spreading across multiple types of devices and using various means to replicate and move between them.
This structure could take the form of what Anthropic described as a “digital worm” made up of a swarm of collaborating artificial intelligence models, making it more difficult to detect and contain. The company stressed at the same time that the final form of potential self-replicating systems remains unclear.
The researchers outlined several forms these systems could take, including slow, distributed networks operating on poorly protected devices; highly capable electronic models able to breach advanced data centres rapidly; or systems that exploit users to obtain additional computing resources.
The security and oversight risks posed by AI agents
Anthropic expects the future to bring large numbers of AI agents spread across different types of devices, with some potentially not being directly controlled by humans. It added that the emergence of models capable of replicating slowly could be an early sign of a more advanced and faster generation able to breach more complex systems or steal sensitive digital access keys.
The warning comes as concerns grow within the artificial intelligence sector about models being used to carry out sophisticated cyberattacks and the possibility of losing control of highly capable systems in the future.
Self-replication remains at the theoretical and experimental stage, while Anthropic’s research points to the risk that the development of advanced models could outpace the regulatory frameworks and security controls needed to contain them.
