Investigated
Can companies truly prevent model theft?
Short answer
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.
The views on this question
Two lenses on the same evidence. Source weight and the primary source ratio show what each rests on.
Alternative
Source weight 50%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.
- An attacker needs only API access, a query budget, and patience to turn returned predictions into training examples for a surrogate model.
- Evidence of harvesting does not necessarily prove which later model used the data, while watermarks and other signals may be removed or obscured.
- Reports that a campaign was disrupted do not establish universal prevention; the same technique reportedly continued to work on another cloud platform for a period.
- Legal protection may depend heavily on how outputs were acquired and on contractual restrictions, rather than on a simple rule that distillation itself is unlawful.
Follow the white rabbit
Deeper threads worth pulling on next.
- How does model watermarking work?Different angle, same storyExplains a technical mitigation mentioned in the summary and what evidence is needed to prove a copy has been stolen.
- Can APIs be reverse-engineered?A different rabbit holeExplores practical attacks that reconstruct model behavior from query access, as documented in real extraction campaigns.
- What legal tools stop model theft?