Kimi K3.1 Pro Scores 64% on the Artificial Analysis Cyber Index

Moonshot AI’s Kimi K3.1 Pro scored 64% on the Artificial Analysis Cyber Index, the highest result among publicly available models, as the company detailed steps to reduce cyber misuse risks.

Moonshot removed cyber-related data from the model’s mid-training corpus and cyber-related tasks from the mix used in its reinforcement-learning environment. The changes were intended to reduce misuse risk ahead of a planned release of the model’s full weights by Nov. 4.

On the Artificial Analysis Cyber Index, K3.1 Pro’s 64% result was 23 percentage points above K3’s score. It also topped DeepSeek V4.1 Flash 1017, which scored 57%. The only higher result in the comparison was 69% for GPT-6 Sol (Daybreak Blue, max), a model available through trusted access.

Moonshot’s xdev-bench-hard benchmark showed a 90.5% pass@3 solve rate for a control version without risk-reduction measures. The final K3.1 Pro model recorded an 81% pass@3 solve rate, a difference of 9.5 percentage points.

K3.1 Pro uses K3’s 2.8 trillion-parameter base. Moonshot used large-scale reinforcement learning to improve the model’s agent capabilities. The model is available through Kimi.com, Kimi Work, Kimi Code and Kimi API.

The complete weights have not yet been released; Moonshot plans to make them available by Nov. 4. The planned weight release follows Moonshot AI’s Kimi K3.1 Pro launch announcement.