China's Moonshot deliberately weakens the hacking skills of its new Kimi model before releasing the weights

What happened? Beijing-based Moonshot AI has launched Kimi K3.1 Pro, which it calls its most capable 3-trillion-parameter-class open-weight model, and says it deliberately cut the model's offensive cyber abilities before releasing it. The company found that cyber-related content in its general training data and reinforcement-learning environments was steadily lifting the model's ability to find and exploit software flaws, so it removed that data from both stages while keeping general coding tasks. The model is live on Kimi's apps and API, and Moonshot says the full weights will be published by 4 November 2026.

The numbers On Moonshot's in-house xdev-bench-hard test, where an AI agent must exploit recently disclosed bugs in software such as the Linux kernel and Node.js, a version trained without the cuts solved 90.5% of tasks; the released K3.1 Pro solved 81.0%. That is still far above the 33.3% scored by the original K3, according to The Next Web, and Moonshot says K3.1 Pro remains behind frontier proprietary models that have been public for at least three months. Reuters notes the results have not been independently verified.

Why it matters "We aim not to release open weights that would extend the absolute frontier of cyber capability," Moonshot wrote, while saying it intends to raise the cyber capability in open models step by step so more teams can audit their own code. Critics point out that once weights are public, the developer can no longer restrict or monitor their use, and Notebookcheck reports that one company has already announced plans to strip K3.1 Pro's safeguards after release. The move comes days after China proposed export controls on AI-driven malware generation and AI-assisted vulnerability-discovery tools, and as OpenAI said it resumed training its most advanced models after a security overhaul.

Sources: Moonshot AI (Kimi blog), Reuters, The Next Web, Notebookcheck