Moonshot deleted the model's hacking training first — its exploit score dropped from 90.5% to 81%

Moonshot says it watched the model's vulnerability-research skills grow as post-training scaled, so it stripped cyber-related data from both mid-training and reinforcement learning for the open-weight release. On its internal exploit suite, the unfiltered build solved 90.5% of tasks; the shipped model solves 81.0% — the company's own measure of what the cut cost.

The company's stated principle: no open-weight release should extend the absolute frontier of offensive cyber capability. The controls apply at training time — the technical post doesn't address whether someone could fine-tune the released weights to put the filtered skills back.