In the tests described there, a 9-billion-parameter model successfully ran inference on eight of nine hardware types Anthropic evaluated. Larger models did better at the task, including China's GLM-5.2, which outperformed the smaller model at replicating itself across devices. Anthropic estimated that if smaller models keep improving at their current pace, they could match the performance of larger models by the end of the year.

To illustrate the risk, Anthropic built a demonstration in which a model ran distributed across 15 interconnected Android TV boxes, treating the devices as a single combined system. The company referred to models that spread across hardware this way as "worms." Anthropic was explicit that the demonstration shows distributed inference is possible across connected devices, not that the devices had been compromised or that the model copied itself onto them without human direction.

Anthropic called on device makers and security teams to tighten protections that make unauthorized software installation and replication harder on consumer hardware.

The post landed days after Anthropic chief executive Dario Amodei called for a slowdown in AI development.