Moonshot AI limits new Kimi model's hacking capabilities as AI cyberattack concerns grow

Moonshot AI reduced the cyberattack capabilities of its open-weight new model Kimi K3.1 Pro when releasing it, and plans to make the weights public by November 4. The company removed cyberattack-related data from its training corpus and reinforcement learning environment while retaining general programming capabilities. The company says K3.1 Pro scored 81.0% on an internal cyber benchmark, lower than the 90.5% of the unreduced version, but still better than the previous generation; the results have not been independently verified.