Moonshot says Kimi K3.1 Pro filtering reduces exploit-solving performance

Moonshot AI says it removed cyber-related training data and reinforcement learning environments from Kimi K3.1 Pro, reducing its performance on the company's internal exploit benchmark. The filtered model solved 81% of tasks, compared with 90.5% for an unfiltered version; general coding tasks remained in training. The Beijing-based startup plans to release the model's open weights by November 4, allowing users to download and run it on their own hardware. Ethereum co-founder Vitalik Buterin endorsed restricting capabilities rather than screening users. The filtering leaves substantial exploit-solving capability: Moonshot says rival open model DeepSeek V4.1-Flash scored 85.7% on the same benchmark.