Moonshot proactively reduces Kimi K3.1 Pro's cyberattack capabilities before open-weight release

Before the open-weight release of Kimi K3.1 Pro, Moonshot AI removed cyberattack-related data from its training corpus and reinforcement learning environment to reduce the model's cyberattack capabilities while trying to preserve its software development capabilities as much as possible; the model weights are scheduled for release on November 4. Moonshot said that on an internal benchmark testing a model's ability to identify and exploit software vulnerabilities, the released version of K3.1 Pro scored 81%, compared with 90.5% for a similar version without the reduction measures, a decrease of 9.5 percentage points; these results come from Moonshot itself and have not been independently verified. The company said K3.1 Pro still performs better than the previous generation model on cyber tasks, and that this is capability limitation rather than elimination. The model also has a 1 million token context window, and the released Highspeed version is said to run 2 to 3 times faster than the standard version.