Kimi K3.1 Pro ranks fourth among 215 AI models; Moonshot AI removed cyberattack training data before the open-weights release

Moonshot AI released Kimi K3.1 Pro on October 23; it ranks fourth among 215 models on the Artificial Analysis Intelligence Index, and before the open-weights release the company proactively removed cyberattack training data. The model scored 81.0% on its in-house cybersecurity evaluation xdev-bench-hard, while an ablation checkpoint without the data removal scored 90.5%; the company says this makes its offensive cyber capability lower than that of frontier proprietary models that have been public for at least three months. The weights are scheduled to be available for download on November 4; routine fine-tuning can in principle partially restore the suppressed capability, and the durability of this reduction remains to be independently tested.