OMM Redaksiyası
Small artificial intelligence models could soon spread among internet-connected devices without human intervention. Anthropic says it may become harder to prevent such systems from replicating on devices such as smart home appliances and Android TV units.
The company tested whether models could identify poorly secured devices, install copies of themselves on them and continue spreading without human assistance.
How close are small models to self-propagation?
The tests showed that small models with fewer than 30 billion parameters can already perform many of the steps required to replicate themselves. However, they have not yet been able to complete the process fully autonomously.
In simplified tests in which some security measures were disabled, GLM-5.2 and more advanced models were able to carry out the full propagation process.
Anthropic stressed that self-replicating AI systems could be particularly difficult to control. The company said there may be no single shutdown mechanism capable of stopping them.
Unlike conventional malware, these systems could adapt to unfamiliar devices without human direction and potentially target processors in data centers.
Security risks in smart devices
The study examined Android TV units, smart home appliances and other internet-connected devices. Anthropic noted that some of these devices have weaker security protections.
The company estimates that if small models continue to improve at their current pace, they could become capable of spreading autonomously by the end of 2026. The ability to transfer their model weights could emerge soon afterward.
Even if these capabilities develop, Anthropic expects the systems to spread relatively slowly by exploiting widely used devices with weak security. The company called for stronger security measures to prevent self-replicating AI from spreading through connected devices.
