Photo Credit: WAM
Artificial intelligence company Anthropic has warned that smaller AI models could be only months away from gaining the ability to autonomously replicate across commercial Internet-of-Things (IoT) devices, raising new cybersecurity and containment concerns.
The company said increasingly capable AI systems could potentially spread across devices such as smart appliances and low-cost Android TV boxes without continuous human assistance. Such systems could prove difficult to contain because there may be no central point from which they can be switched off.
Unlike conventional malware, autonomous AI agents could potentially adapt to unfamiliar hardware, identify ways to operate on different devices and seek additional computing resources, including more powerful data centre processors.
Anthropic examined the possibility through experiments designed to test whether AI models could locate vulnerable devices, transfer themselves to those systems and continue spreading independently.
The research found that models with 30 billion parameters or fewer can already perform many of the individual tasks required for autonomous replication. However, they still encounter difficulties when attempting to complete the entire process consistently from beginning to end.
In simplified testing environments where some cybersecurity barriers were removed, GLM-5.2 and more capable models demonstrated end-to-end replication abilities. Anthropic said that if smaller models continue improving at their historical pace, they could reach similar capability levels by the end of 2026.
The company also indicated that small weight-based AI replicators, which carry or deploy model weights directly onto compromised hardware, could become technically possible soon afterwards.
Anthropic said improving cybersecurity across internet-connected devices will be increasingly important as autonomous AI capabilities advance. IoT products such as smart televisions, household appliances and other connected devices could represent attractive targets because many have comparatively weak security protections.
The company noted that early autonomous replicators would likely spread relatively slowly and depend heavily on widely available and less-secure computing resources. However, continued improvements in smaller AI models could increase the security risks associated with autonomous replication over time.
