Moonshot AI filtered cyber-related data from the training corpus of Kimi K3.1 Pro to reduce its ability to exploit software vulnerabilities while preserving its coding skills. On Moonshot’s internal exploitation benchmark, the filtered model scores 81.0%, compared with 90.5% for an unfiltered version and 33.3% for the preceding Kimi K3. With 2.8 trillion total parameters, including 104 billion active parameters, and a one-million-token context window, the model scores 65.7% on Terminal-Bench 4.0, comparable to Anthropic’s Claude Opus 5.5. Moonshot plans to release the full model weights on November 4, 2026, allowing researchers and security teams to independently evaluate the effectiveness of these safeguards.
Source: Read the original article
