Anthropic: AI Models May Soon Begin Self-Copying

Anthropic stated that its experiments suggest small AI models may soon be capable of self-replicating on commercial Internet of Things (IoT) devices, such as smart ovens and Android-based televisions from unknown brands, within a few months.

In its experiments, Anthropic investigated whether AI models could identify devices with security vulnerabilities, replicate themselves onto those devices, and continue spreading without human intervention.

The company found that models with fewer than 30 billion parameters could complete the multiple steps required for self-replication. However, it noted that these AI models struggled to complete the entire sequence from start to finish.

In its simplified tests, where certain security measures were bypassed, Anthropic found that the Llama 3.2 model and stronger models were capable of full self-replication from start to finish.

Anthropic warned that self-replicating models would be difficult to stop due to the lack of a central control point to disable them. Unlike traditional malware, they can autonomously detect unfamiliar devices and target data center processors.

The company focused its research on IoT devices such as Android-based televisions, smart home appliances, and other internet-connected devices, as these are typically insufficiently protected.

The leading AI company also warned that if small models continue to improve at their unprecedented pace, they will be able to self-replicate independently by the end of this year, with weight-based small replication models following shortly after.

Anthropic also noted that these autonomous AI systems are likely to replicate slowly, relying on widely available computing resources and lacking sufficient security. It added that systems should be strengthened to make the spread of these AI agents more difficult.