Moonshot trims Kimi’s hacking skills and still tops the open cyber table

Moonshot AI has done something unusual for a model maker on launch day: it has published evidence that it made its product worse. The Beijing laboratory released Kimi K3.1 Pro on 23 October and says it removed cyber-related material from the model’s mid-training data and cyber tasks from its reinforcement-learning environments before the open weights ship, due by 4 November. On Moonshot’s in-house exploit test, the released model solved 81.0% of problems within three attempts, against 90.5% for a build trained with the material left in.

This article continues for subscribers. Subscribe to read the full analysis.