Unauthorized distillation
Using a model's outputs without permission to train another model that reproduces its capabilities.
Definition
The report describes actors who used API access or stolen keys to generate large volumes of outputs, then trained their own models on them.
This effectively transfers the model's capabilities without paying for or being authorized to do so.
{esc(t(CURRENT_LANG, "not_confused_with"))}
- Not legitimate distillation with the model owner's permission.
- Not directly stealing model weights.
{esc(t(CURRENT_LANG, "example"))}
See the supply chain intrusion case.