The deciding number is 8 of 8 nodes for GLM-5.2. That supports a lab result, not a deployed threat.

Replication is a chain, not a single capability. A model that can set up inference on unfamiliar hardware is not yet a worm that finds targets, gains execution, and keeps spreading.

The primary page leaves the exploit stages out. The buyer or operator should treat the claim as a capability ceiling, not a field incident.

Fifteen $55 Android TV boxes run Gemma4-26B-A4B at 0.39 tokens per second. That is slow, but enough to show cheap nodes can host inference.

The constraint is firmware, update cadence, and network isolation. Those details matter for the fab and the device maker.

The page gives a doubling time of days. It does not give a measured infection rate, a failure rate under patching, or a cost per compromised node.

Those numbers decide whether the threat is a lab curiosity or a supply-chain problem.

The wire turns a lab capability into a present-tense threat. The primary page only proves that small models can set up inference and that GLM-5.2 can chain the steps in simplified tests.

The page is the safer record, and the wire's present-tense threat overreaches.