The report says the company removed cyber-related material and vulnerability-research tasks from both mid-training data and reinforcement-learning environments.
The reported aim was to avoid releasing weights that would extend the “absolute frontier” of cyber capability. Moonshot reportedly retained general-purpose coding capabilities while reducing training exposure related to finding and exploiting vulnerabilities.
The article cites an internal controlled-environment comparison: a version without the risk-reduction measures scored 90.5% on finding and exploiting software vulnerabilities, while the publicly released Kimi K3.1 Pro scored 81.0%.
The supplied source does not provide Moonshot’s underlying evaluation methodology, task set, model weights, licensing terms, release date, or an independent verification of either score. It also does not establish how the reported reduction affects real-world misuse or whether downstream users can alter the model’s behavior.