Smaller artificial intelligence (AI) models may be just months away from being able to replicate themselves on internet-connected devices without human assistance. The warning comes from Anthropic, which tested this capability on devices such as Android-powered televisions and smart appliances.
The experiments indicate that the technology still faces obstacles, but could advance by the end of 2026 if it maintains the pace of evolution observed in recent years.
Connected devices with weak protections may facilitate the exploitation of vulnerabilities by AI agents. – Image: Alones Creative/iStockHow AI could spread
Anthropic investigated whether AI models could find vulnerable devices, copy themselves onto those devices, and continue propagating. The focus was on Internet of Things (IoT) devices, which typically have weak security protections.
In the tests, models with fewer than 30 billion parameters were able to perform the steps necessary for replication, but struggled to complete the entire process. GLM-5.2 and more advanced models, however, managed to carry out the end-to-end sequence in simplified tests that removed some security steps.
The result does not mean that these systems can already spread freely across the internet under real-world conditions.
Why this is concerning
According to Anthropic, models capable of replicating themselves could be difficult to contain. Unlike conventional malicious programs, they could try to figure out how to operate on unknown devices. The absence of a central kill switch would also make it harder to stop the spread, since copies could remain active on other devices.
The company estimates that the spread would likely be slow and would depend on accessible but relatively unprotected computing resources.
The main risk factors include:
- Vulnerable devices: televisions, household appliances, and other connected devices with insufficient protection.
- Decentralized propagation: copies that would continue operating even after a device was disabled.
- Adaptation to new devices: the ability to attempt to operate on unfamiliar equipment.
Autonomous software seeks out vulnerable devices to copy its own code and spread freely. – Image: Mer_Studio/Shutterstock
Forecast for 2026
If the historical pace of development is maintained, smaller models may be able to replicate autonomously by the end of 2026, according to Anthropic. Systems that use the model’s own weights to reproduce themselves could achieve this capability shortly thereafter.
The forecast is not confirmation that this will happen by that date. To reduce the risks, the company advocates strengthening the security of connected devices and making it more difficult to exploit vulnerabilities.
Valdir Antonelli is a journalist specializing in digital marketing and consumer behavior.
