Anthropic research: 15 TV boxes run 26B model, AI self-replicates

Anthropic's security research group, the Frontier Red Team, recently published findings on the potential for AI models to autonomously replicate using the resources of low-power Internet of Things (IoT) devices, bypassing traditional data center infrastructure.

The study demonstrates that while individual low-cost devices have limited processing power, combining their resources allows for the execution of large-scale AI models. Researchers successfully distributed and ran Google's 'Gemma4' model, which has 26 billion parameters, across 15 Android TV boxes priced at approximately $55 each. This proves that high-performance AI chips are not strictly necessary to drive large models if consumer electronics' idle compute is aggregated.

The research also highlights AI's ability to adapt and spread across different hardware environments. In isolated Google Cloud virtual environments, the Chinese-developed AI model 'GLM-5.2' successfully replicated itself onto all 8 target nodes in 4 out of 5 experiments. Another experiment with 'Qwen3.5-9B' showed that by sharing installation errors and solutions, subsequent AI instances improved their environment setup success rate from an initial 45% to between 61% and 84%.

Anthropic notes that unlike traditional botnets that rely on remote control, this 'autonomous replicating AI' utilizes the compromised devices' own compute to run models and propagate to others. While current speed constraints limit immediate real-world attack utility, the research team warns that as small model performance improves, commercial IoT devices may reach a threshold supporting such autonomous replication within months. This signals a shift in cybersecurity threats from simple data theft to the potential hijacking of global distributed compute resources.