Moonshot trims Kimi K3.1 Pro’s hacking score before open-weight release
Moonshot AI says it removed cyber-related material from Kimi K3.1 Pro’s training and reinforcement-learning environments while retaining general coding skills. The company reports that the model scored 81.0% on its internal vulnerability-finding test, against 90.5% for an unfiltered version; the results have not been independently verified. Moonshot plans to publish the model’s weights by 4 November.
Bottom line — The 4 November release will put Moonshot’s safeguards to a wider test, but distributed weights are harder for their developer to monitor or restrict.
Go deeper 5
-
Reuters reports that Moonshot’s controlled test involved finding and exploiting software vulnerabilities.
-
Moonshot says the filtered model still outperformed its predecessor on the cyber test.
-
Reuters notes that open-weight models can help organisations inspect their own software, including older systems.
-
Once weights circulate, Reuters says, Moonshot has less ability to monitor use, restrict access or update safeguards.
-
China has proposed draft export controls on AI-driven malware generation and mutation, and restrictions on AI-assisted vulnerability discovery; the proposals remain open for public comment.