Moonshot says training restrictions reduce Kimi K3.1 Pro's hacking performance

Moonshot AI says it reduced Kimi K3.1 Pro's offensive cybersecurity capabilities by removing cyber-related data from training and reinforcement-learning environments while retaining general coding capabilities. The company introduced the model on October 24. In Moonshot's controlled vulnerability-finding and exploitation benchmark, the restricted model scored 81%, compared with 90.5% for a similar version without the measures. Those internal results have not been independently verified. Moonshot plans to release the model's weights on November 4, letting developers download, run and modify it themselves. Once those files are distributed, the company has less control over who uses the model and whether its safeguards remain effective.