N環球 
Moonshot curbs Kimi K3.1 Pro hacking ability, but its own test still shows 81% exploit score
Moonshot AI said on Oct 24 it stripped cyber-related data from Kimi K3.1 Pro training while keeping general coding skills, with weights still due by Nov 4; its unpublished in-house exploit test still scores 81.0%, versus 90.5% unfiltered.
Why it matters
Beijing-based Moonshot AI said on Saturday, Oct 24, that it has reduced the cyber-offensive capabilities of its newest model, Kimi K3.1 Pro, Reuters reported, as fears grow that increasingly autonomous AI could make hacking easier and cheaper.
Moonshot said it removed cyber-related data from the training corpus and reinforcement-learning environments used to develop the model, while retaining general coding skills. The company said: We aim not to release open weights that would extend the absolute frontier of cyber capability. The full weights are still due by Nov 4.
The cut is partial
The cut is not a wipeout. On Moonshot's unpublished in-house exploit test, xdev-bench-hard, K3.1 Pro still solved 81.0% of problems, against 90.5% for an unfiltered build. The company has not said whether fine-tuning could restore the removed capability, and open weights cannot be recalled once downloaded.
The backdrop is China's proposed technology export controls, announced Oct 16: a draft that would ban exports of AI-driven malware-generation and mutation technology and restrict AI-assisted vulnerability-discovery tools, with public comments open until Nov 15. Moonshot argues open-weight models can also help more organisations inspect and secure their own software, including older systems that receive little attention.
K3.1 Pro, released Oct 23, is described by the company as its most capable 3-trillion-parameter-class open-weight model. Its vendor-reported Terminal-Bench 4.0 score is 65.7%, above GPT-6 Astra at 58.2% and just below Claude Opus 5.5 at 66.4% — vendor figures, not yet independently verified.
💬 Comments
No comments yet. Be the first!