Anthropic tests how AI could replicate across connected devices.

Anthropic’s Frontier Red Team studied inference setup and replication on connected devices. Its report suggests smaller models may be nearing the required capabilities.

What the researchers tested

The team studied inference setup on unfamiliar hardware and replication in a controlled fleet. A nine-billion-parameter model set up inference on eight of nine tested hardware types. Larger, more capable models performed better at chaining the steps together in simplified evaluations.

The limit that matters

The replication evaluation simplified the initial-access challenge. The researchers describe its results as an upper bound on real-world success. A demonstrated capability in that environment does not establish that a model is spreading independently on the public internet.

Why it deserves attention

The report highlights connected consumer devices as possible computing resources for future AI-enabled malware. Its timeline is an estimate based on capability trends, rather than a confirmed prediction. The practical distinction is between a research finding, a plausible threat and an observed incident.