Moonshot restricts hacking ability of new Kimi model

Moonshot said it removed cybersecurity-related data from the training datasets, without impairing the model's general coding ability. The Beijing-based company said it aims not to release open-weight models that push the absolute frontier of cyber capabilities. It calls Kimi K3.1 Pro its most powerful open-weight model, trained with three trillion parameters, and plans to publish its weights by 4 November, which will let outside developers run and modify it on their own systems. Last week, the Chinese government proposed export controls that would ban exporting AI technology for creating and altering malware and restrict AI tools for finding vulnerabilities, Handelsblatt reports.