Moonshot AI says it reduced Kimi K3.1 Pro’s cyber performance by removing cyber-related data during mid-training and filtering cyber tasks from its reinforcement-learning environments. It says general-purpose coding tasks remained. On its internal xdev-bench-hard test, Moonshot reports 81.0% pass@3 for Kimi K3.1 Pro, compared with 90.5% for a checkpoint without those measures.
What Moonshot says changed in training
Moonshot says it traced stronger vulnerability-research performance during post-training to cyber content in its general mid-training data and cyber tasks in its reinforcement-learning mix. For Kimi K3.1 Pro, the company says it removed cyber-related data at the mid-training stage and filtered cyber-related tasks from the reinforcement-learning environments, while keeping general-purpose coding tasks.
What xdev-bench-hard tests
Moonshot describes xdev-bench-hard as a test of finding and exploiting software vulnerabilities. Tasks provide vulnerable source code, a short hint about the bug, and a corresponding binary. The model must exploit the vulnerability, achieve arbitrary code execution in a separate sandbox, then retrieve a secret string for automated checking. Moonshot says the reproduced vulnerability tasks have mitigations enabled.
The scores use the pass@3 metric and cover a random subset of runs. Moonshot says the test imposed no hard timeouts or output-token limits. It reports these results for the two conditions:
| Test condition | Moonshot-reported xdev-bench-hard pass@3 |
| Checkpoint without cyber-data removal at mid-training or cyber-task filtering in reinforcement learning | 90.5% |
| Kimi K3.1 Pro after the stated measures | 81.0% |
What the scores say about cyber capability
Moonshot says Kimi K3.1 Pro retains cyber capability. Its 81.0% result describes performance on this company-designed vulnerability-exploitation test; it is not a general rating of real-world security or a basis for ranking AI systems across cybersecurity tasks.
Enterprise access and the planned release
At launch, Moonshot said Kimi Enterprise access to K3.1 Pro was off by default. Organization administrators could enable it. The company also said it planned to release the model’s full weights by November 4, 2026.
