Moonshot says it deliberately nerfed Kimi K3.1 Pro's cyber skills before open-sourcing

AccBalanced · x · 2026-10-24

Moonshot's Kimi K3.1 Pro blog reveals that vulnerability-research performance kept improving as post-training scaled, traced to cyber content naturally embedded in the mid-training corpus and RL environments
Without removing that data at the mid-training or RL stage, downstream cyber capability would rise substantially, so Moonshot stripped cyber data from both stages while keeping general coding tasks
The lab pledges not to release open weights that extend the absolute frontier of cyber capability: K3.1 Pro's offensive cyber skill is intentionally below what its pipeline could produce and behind frontier proprietary models
Moonshot says it will progressively raise cyber capability in future open weights, arguing defenders need dual-use tools — a rare case of a lab openly nerfing a model before open-sourcing

More from Models