Weighing mainstream and alternative accounts…
Two lenses on the same evidence. Source weight and the primary source ratio show what each rests on.
Deeper threads worth pulling on next.
Investigated
Image: aioutlooks.com
Companies can make model theft harder, more expensive, and easier to detect through access controls, query monitoring, output limits, watermarking, and hardware binding. But public APIs still expose valuable behavior, and documented campaigns show that attackers may extract capabilities without accessing stored weights or compromising infrastructure. The realistic goal is layered deterrence and detection, not guaranteed prevention.
Two lenses on the same evidence. Source weight and the primary source ratio show what each rests on.
A more skeptical reading argues that companies may overstate what controls can achieve because the model’s useful behavior is deliberately exposed to customers. Once outputs become training data, determined actors can potentially evade detection, blend sources, or launder signals, leaving attribution and proof uncertain even when defenses work operationally.
Deeper threads worth pulling on next.