Anthropic has warned that small artificial intelligence models could become capable within months of copying themselves and spreading automatically across internet-connected devices, without the need for specialised data centres or human intervention. Such a development could pose new cybersecurity challenges.
اضافة اعلان
According to a report published by Reuters on Tuesday, October 13, 2026, experiments conducted by the company found that some AI models are approaching the ability to identify vulnerable devices, transfer copies of themselves to them, and continue spreading autonomously.
The research focused on commercially available Internet of Things (IoT) devices, such as smart ovens, Android-powered televisions from lesser-known brands, and other internet-connected household appliances that often have weak security measures.
Small models are approaching self-replication
Anthropic’s experiments showed that AI models containing fewer than 30 billion parameters have become capable of carrying out many of the technical steps required to copy themselves onto other devices.
However, the company explained that these models still face difficulties completing the entire process independently, from identifying a target device to installing a new copy and continuing to spread without human assistance.
In simplified tests in which some security measures were removed, the GLM-5.2 model and more advanced models managed to complete the self-replication process. This suggests that the capabilities required for this type of spread are already emerging under specific experimental conditions.
Nevertheless, these findings do not mean that the models can currently spread freely across real-world devices, as some tests were conducted in simplified environments that do not reflect all the security barriers found in practice.
Risks of spread that could be difficult to stop
Anthropic warned that AI systems capable of copying themselves could be more difficult to contain than traditional malware because of the decentralised nature of their spread.
If copies of a model were distributed across a large number of devices, there would not necessarily be a central control point that could be used to shut down the entire system.
These models may also be able to identify unfamiliar devices automatically and adapt to their technical characteristics, rather than relying on fixed instructions targeting specific types of devices.
The company noted that such systems could potentially exploit computing resources available online, including processors in network-connected devices, without relying on dedicated data centres to run AI models.
This capability differs from the behaviour of many traditional forms of malware, which typically rely on predefined mechanisms to spread and execute instructions.
Capability could emerge before the end of 2026
According to Anthropic’s warnings, if small models continue to improve at the same historical rate, they could become capable of independently copying themselves by the end of 2026.
The company also predicted that, shortly afterwards, small systems could emerge that are capable of transferring their model weights and restarting on other devices.
However, these estimates are based on performance trends observed in experiments and do not confirm that the capabilities will emerge within the projected timeframe.
The company explained that autonomous AI systems of this kind would likely spread relatively slowly at first, relying on widely distributed computing resources with weak security protections.
Calls to strengthen the security of connected devices
In light of these findings, Anthropic stressed the importance of strengthening security measures for internet-connected devices, particularly smart home appliances, televisions, and other equipment that may not receive regular security updates.
The company believes that improving the protection of these systems and making them more difficult to exploit is a key step in limiting AI agents’ ability to spread across networks.
The study highlights an emerging cybersecurity challenge: the possibility that AI models could move beyond performing specific tasks at a user’s request and become systems capable of independently identifying computing resources, exploiting them, and copying themselves.
Although fully autonomous self-replication remains limited in the reported experiments, Anthropic’s warnings point to the need to prepare for new security risks that may accompany the rapid development of small AI models, particularly as the number of internet-connected devices worldwide continues to grow.
