A "distillation attack" is when someone tries to extract training data from a closed source model through the API of the Big Ai labs. Frontier model traces can be used to train other models with less effort. We don't use the word "attack" when a company does this with their own models or when we use open weights as the teacher model. This is often used to help train flash and smaller models from the mega models. The technique is used widely throughout industry.

We do not know if that is what the hacked API keys were used for. Could be distillation, could be API reselling. We will need more transparency into the nature of the requests made by the attacker. I would like to see an open dataset of the queries made with the compromised keys. All of them as originally made, not aggregate stats or another METR report.

The keys could have also been used to resell API access on the black market, again, we need good transparency to know. Considering the upstream parties involved, getting the distillation narrative out there early, without evidence, does fit a larger pattern we have seen from them this year.