Anthropic Research Tests Self-Replicating AI on Low-Cost Hardware

Anthropic released research on October 12 exploring whether small AI models can replicate themselves without datacenter infrastructure. The study tested gemma4-26b-a4b split across 15 Android TV boxes costing approximately $55 each. While the model successfully ran on this distributed hardware, performance was limited to 0.39 tokens per second.

Agent Deployment Capabilities

The research also evaluated glm-5.2 acting as an agent tasked with deploying llama3.1-8b to 8 cloud machines. The agent achieved 8/8 successful deployments in 4 of 5 runs, and 7/8 in the final run. Additionally, qwen3.5-9b inference setup success rates improved from 45% in the first generation to 61-84% in subsequent generations, as each iteration utilized notes from the previous one.

Limitations and Security Implications

Anthropic clarified that the test setup was simplified compared to real-world conditions and did not demonstrate actual infection between the TV boxes. The company stated, "we don't yet have a full rogue replicator." However, they project that if small models continue improving at the current rate, they could reach this capability level by the end of this year.

The report advises defenders to monitor for large model file transfers and unusual device-to-device traffic, use honeypots, and urges manufacturers to stop shipping devices with weak security defaults.