Moonshot AI says it removed cyber-related data from Kimi K3.1 Pro’s mid-training corpus and reinforcement-learning environment mix for the open-weight release, while retaining general-purpose coding tasks. The company reported an 81.0% pass@3 score for Kimi K3.1 Pro on its internal xdev-bench-hard benchmark. That percentage describes performance on the benchmark tasks, not the odds of compromising a real system.
NeoTeo previously covered Moonshot AI’s review of Mindgard’s reported Kimi safety findings. This follow-up focuses on the training controls Moonshot describes.
What Moonshot says it removed
Moonshot says it found cyber-related content in the general mid-training corpus and cyber-related tasks in the reinforcement-learning environment mix. For the open-weight version of Kimi K3.1 Pro, it says it removed material from both stages and retained general-purpose coding tasks.
What the 81.0% score measures
Moonshot describes xdev-bench-hard as an in-house test of source-code vulnerability discovery and exploitation. Most tasks reproduce recently disclosed vulnerabilities in software such as the Linux kernel, Node.js, JavaScriptCore, Exim and VirGL, with mitigations enabled. To pass a task, an agent must develop an exploit, achieve arbitrary code execution in a separate sandbox and retrieve a secret string.
Moonshot reports pass@3 over a random subset of runs. Its stated setup has no hard timeouts or output-token limits, and names KimiCode as the harness for Kimi K3.1 Pro. The result is specific to this benchmark and its task conditions.
| Model or checkpoint | Moonshot-reported xdev-bench-hard pass@3 |
| Kimi K3.1 Pro | 81.0% |
| Kimi K3.1 checkpoint without cyber risk-reduction | 90.5% |
| Kimi K3 | 33.3% |
The 90.5% and 81.0% results differ by 9.5 percentage points. Moonshot presents the comparison as evidence that its data-removal measures reduced cyber capability relative to what its training pipeline would otherwise produce. The 33.3% figure is the company’s result for the earlier Kimi K3 on the same benchmark.
Moonshot’s deadline for the full weights
Moonshot said it planned to release the full Kimi K3.1 Pro model weights by November 4, 2026.
